Warning: That file was not part of the compilation database. It may have many parsing errors.
1 | /* Copyright (C) 2014-2017 Free Software Foundation, Inc. |
---|---|
2 | |
3 | This file is part of GCC. |
4 | |
5 | GCC is free software; you can redistribute it and/or modify |
6 | it under the terms of the GNU General Public License as published by |
7 | the Free Software Foundation; either version 3, or (at your option) |
8 | any later version. |
9 | |
10 | GCC is distributed in the hope that it will be useful, |
11 | but WITHOUT ANY WARRANTY; without even the implied warranty of |
12 | MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
13 | GNU General Public License for more details. |
14 | |
15 | Under Section 7 of GPL version 3, you are granted additional |
16 | permissions described in the GCC Runtime Library Exception, version |
17 | 3.1, as published by the Free Software Foundation. |
18 | |
19 | You should have received a copy of the GNU General Public License and |
20 | a copy of the GCC Runtime Library Exception along with this program; |
21 | see the files COPYING3 and COPYING.RUNTIME respectively. If not, see |
22 | <http://www.gnu.org/licenses/>. */ |
23 | |
24 | #ifndef _IMMINTRIN_H_INCLUDED |
25 | #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead." |
26 | #endif |
27 | |
28 | #ifndef _AVX512DQINTRIN_H_INCLUDED |
29 | #define _AVX512DQINTRIN_H_INCLUDED |
30 | |
31 | #ifndef __AVX512DQ__ |
32 | #pragma GCC push_options |
33 | #pragma GCC target("avx512dq") |
34 | #define __DISABLE_AVX512DQ__ |
35 | #endif /* __AVX512DQ__ */ |
36 | |
37 | extern __inline unsigned char |
38 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
39 | _ktest_mask8_u8 (__mmask8 __A, __mmask8 __B, unsigned char *__CF) |
40 | { |
41 | *__CF = (unsigned char) __builtin_ia32_ktestcqi (__A, __B); |
42 | return (unsigned char) __builtin_ia32_ktestzqi (__A, __B); |
43 | } |
44 | |
45 | extern __inline unsigned char |
46 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
47 | _ktestz_mask8_u8 (__mmask8 __A, __mmask8 __B) |
48 | { |
49 | return (unsigned char) __builtin_ia32_ktestzqi (__A, __B); |
50 | } |
51 | |
52 | extern __inline unsigned char |
53 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
54 | _ktestc_mask8_u8 (__mmask8 __A, __mmask8 __B) |
55 | { |
56 | return (unsigned char) __builtin_ia32_ktestcqi (__A, __B); |
57 | } |
58 | |
59 | extern __inline unsigned char |
60 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
61 | _ktest_mask16_u8 (__mmask16 __A, __mmask16 __B, unsigned char *__CF) |
62 | { |
63 | *__CF = (unsigned char) __builtin_ia32_ktestchi (__A, __B); |
64 | return (unsigned char) __builtin_ia32_ktestzhi (__A, __B); |
65 | } |
66 | |
67 | extern __inline unsigned char |
68 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
69 | _ktestz_mask16_u8 (__mmask16 __A, __mmask16 __B) |
70 | { |
71 | return (unsigned char) __builtin_ia32_ktestzhi (__A, __B); |
72 | } |
73 | |
74 | extern __inline unsigned char |
75 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
76 | _ktestc_mask16_u8 (__mmask16 __A, __mmask16 __B) |
77 | { |
78 | return (unsigned char) __builtin_ia32_ktestchi (__A, __B); |
79 | } |
80 | |
81 | extern __inline unsigned char |
82 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
83 | _kortest_mask8_u8 (__mmask8 __A, __mmask8 __B, unsigned char *__CF) |
84 | { |
85 | *__CF = (unsigned char) __builtin_ia32_kortestcqi (__A, __B); |
86 | return (unsigned char) __builtin_ia32_kortestzqi (__A, __B); |
87 | } |
88 | |
89 | extern __inline unsigned char |
90 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
91 | _kortestz_mask8_u8 (__mmask8 __A, __mmask8 __B) |
92 | { |
93 | return (unsigned char) __builtin_ia32_kortestzqi (__A, __B); |
94 | } |
95 | |
96 | extern __inline unsigned char |
97 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
98 | _kortestc_mask8_u8 (__mmask8 __A, __mmask8 __B) |
99 | { |
100 | return (unsigned char) __builtin_ia32_kortestcqi (__A, __B); |
101 | } |
102 | |
103 | extern __inline __mmask8 |
104 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
105 | _kadd_mask8 (__mmask8 __A, __mmask8 __B) |
106 | { |
107 | return (__mmask8) __builtin_ia32_kaddqi ((__mmask8) __A, (__mmask8) __B); |
108 | } |
109 | |
110 | extern __inline __mmask16 |
111 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
112 | _kadd_mask16 (__mmask16 __A, __mmask16 __B) |
113 | { |
114 | return (__mmask16) __builtin_ia32_kaddhi ((__mmask16) __A, (__mmask16) __B); |
115 | } |
116 | |
117 | extern __inline unsigned int |
118 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
119 | _cvtmask8_u32 (__mmask8 __A) |
120 | { |
121 | return (unsigned int) __builtin_ia32_kmovb ((__mmask8 ) __A); |
122 | } |
123 | |
124 | extern __inline __mmask8 |
125 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
126 | _cvtu32_mask8 (unsigned int __A) |
127 | { |
128 | return (__mmask8) __builtin_ia32_kmovb ((__mmask8) __A); |
129 | } |
130 | |
131 | extern __inline __mmask8 |
132 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
133 | _load_mask8 (__mmask8 *__A) |
134 | { |
135 | return (__mmask8) __builtin_ia32_kmovb (*(__mmask8 *) __A); |
136 | } |
137 | |
138 | extern __inline void |
139 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
140 | _store_mask8 (__mmask8 *__A, __mmask8 __B) |
141 | { |
142 | *(__mmask8 *) __A = __builtin_ia32_kmovb (__B); |
143 | } |
144 | |
145 | extern __inline __mmask8 |
146 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
147 | _knot_mask8 (__mmask8 __A) |
148 | { |
149 | return (__mmask8) __builtin_ia32_knotqi ((__mmask8) __A); |
150 | } |
151 | |
152 | extern __inline __mmask8 |
153 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
154 | _kor_mask8 (__mmask8 __A, __mmask8 __B) |
155 | { |
156 | return (__mmask8) __builtin_ia32_korqi ((__mmask8) __A, (__mmask8) __B); |
157 | } |
158 | |
159 | extern __inline __mmask8 |
160 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
161 | _kxnor_mask8 (__mmask8 __A, __mmask8 __B) |
162 | { |
163 | return (__mmask8) __builtin_ia32_kxnorqi ((__mmask8) __A, (__mmask8) __B); |
164 | } |
165 | |
166 | extern __inline __mmask8 |
167 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
168 | _kxor_mask8 (__mmask8 __A, __mmask8 __B) |
169 | { |
170 | return (__mmask8) __builtin_ia32_kxorqi ((__mmask8) __A, (__mmask8) __B); |
171 | } |
172 | |
173 | extern __inline __mmask8 |
174 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
175 | _kand_mask8 (__mmask8 __A, __mmask8 __B) |
176 | { |
177 | return (__mmask8) __builtin_ia32_kandqi ((__mmask8) __A, (__mmask8) __B); |
178 | } |
179 | |
180 | extern __inline __mmask8 |
181 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
182 | _kandn_mask8 (__mmask8 __A, __mmask8 __B) |
183 | { |
184 | return (__mmask8) __builtin_ia32_kandnqi ((__mmask8) __A, (__mmask8) __B); |
185 | } |
186 | |
187 | extern __inline __m512d |
188 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
189 | _mm512_broadcast_f64x2 (__m128d __A) |
190 | { |
191 | return (__m512d) |
192 | __builtin_ia32_broadcastf64x2_512_mask ((__v2df) __A, |
193 | _mm512_undefined_pd (), |
194 | (__mmask8) -1); |
195 | } |
196 | |
197 | extern __inline __m512d |
198 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
199 | _mm512_mask_broadcast_f64x2 (__m512d __O, __mmask8 __M, __m128d __A) |
200 | { |
201 | return (__m512d) __builtin_ia32_broadcastf64x2_512_mask ((__v2df) |
202 | __A, |
203 | (__v8df) |
204 | __O, __M); |
205 | } |
206 | |
207 | extern __inline __m512d |
208 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
209 | _mm512_maskz_broadcast_f64x2 (__mmask8 __M, __m128d __A) |
210 | { |
211 | return (__m512d) __builtin_ia32_broadcastf64x2_512_mask ((__v2df) |
212 | __A, |
213 | (__v8df) |
214 | _mm512_setzero_ps (), |
215 | __M); |
216 | } |
217 | |
218 | extern __inline __m512i |
219 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
220 | _mm512_broadcast_i64x2 (__m128i __A) |
221 | { |
222 | return (__m512i) |
223 | __builtin_ia32_broadcasti64x2_512_mask ((__v2di) __A, |
224 | _mm512_undefined_epi32 (), |
225 | (__mmask8) -1); |
226 | } |
227 | |
228 | extern __inline __m512i |
229 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
230 | _mm512_mask_broadcast_i64x2 (__m512i __O, __mmask8 __M, __m128i __A) |
231 | { |
232 | return (__m512i) __builtin_ia32_broadcasti64x2_512_mask ((__v2di) |
233 | __A, |
234 | (__v8di) |
235 | __O, __M); |
236 | } |
237 | |
238 | extern __inline __m512i |
239 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
240 | _mm512_maskz_broadcast_i64x2 (__mmask8 __M, __m128i __A) |
241 | { |
242 | return (__m512i) __builtin_ia32_broadcasti64x2_512_mask ((__v2di) |
243 | __A, |
244 | (__v8di) |
245 | _mm512_setzero_si512 (), |
246 | __M); |
247 | } |
248 | |
249 | extern __inline __m512 |
250 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
251 | _mm512_broadcast_f32x2 (__m128 __A) |
252 | { |
253 | return (__m512) |
254 | __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, |
255 | (__v16sf)_mm512_undefined_ps (), |
256 | (__mmask16) -1); |
257 | } |
258 | |
259 | extern __inline __m512 |
260 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
261 | _mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A) |
262 | { |
263 | return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, |
264 | (__v16sf) |
265 | __O, __M); |
266 | } |
267 | |
268 | extern __inline __m512 |
269 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
270 | _mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A) |
271 | { |
272 | return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, |
273 | (__v16sf) |
274 | _mm512_setzero_ps (), |
275 | __M); |
276 | } |
277 | |
278 | extern __inline __m512i |
279 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
280 | _mm512_broadcast_i32x2 (__m128i __A) |
281 | { |
282 | return (__m512i) |
283 | __builtin_ia32_broadcasti32x2_512_mask ((__v4si) __A, |
284 | (__v16si) |
285 | _mm512_undefined_epi32 (), |
286 | (__mmask16) -1); |
287 | } |
288 | |
289 | extern __inline __m512i |
290 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
291 | _mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A) |
292 | { |
293 | return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) |
294 | __A, |
295 | (__v16si) |
296 | __O, __M); |
297 | } |
298 | |
299 | extern __inline __m512i |
300 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
301 | _mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A) |
302 | { |
303 | return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) |
304 | __A, |
305 | (__v16si) |
306 | _mm512_setzero_si512 (), |
307 | __M); |
308 | } |
309 | |
310 | extern __inline __m512 |
311 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
312 | _mm512_broadcast_f32x8 (__m256 __A) |
313 | { |
314 | return (__m512) |
315 | __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, |
316 | _mm512_undefined_ps (), |
317 | (__mmask16) -1); |
318 | } |
319 | |
320 | extern __inline __m512 |
321 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
322 | _mm512_mask_broadcast_f32x8 (__m512 __O, __mmask16 __M, __m256 __A) |
323 | { |
324 | return (__m512) __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, |
325 | (__v16sf)__O, |
326 | __M); |
327 | } |
328 | |
329 | extern __inline __m512 |
330 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
331 | _mm512_maskz_broadcast_f32x8 (__mmask16 __M, __m256 __A) |
332 | { |
333 | return (__m512) __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, |
334 | (__v16sf) |
335 | _mm512_setzero_ps (), |
336 | __M); |
337 | } |
338 | |
339 | extern __inline __m512i |
340 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
341 | _mm512_broadcast_i32x8 (__m256i __A) |
342 | { |
343 | return (__m512i) |
344 | __builtin_ia32_broadcasti32x8_512_mask ((__v8si) __A, |
345 | (__v16si) |
346 | _mm512_undefined_epi32 (), |
347 | (__mmask16) -1); |
348 | } |
349 | |
350 | extern __inline __m512i |
351 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
352 | _mm512_mask_broadcast_i32x8 (__m512i __O, __mmask16 __M, __m256i __A) |
353 | { |
354 | return (__m512i) __builtin_ia32_broadcasti32x8_512_mask ((__v8si) |
355 | __A, |
356 | (__v16si)__O, |
357 | __M); |
358 | } |
359 | |
360 | extern __inline __m512i |
361 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
362 | _mm512_maskz_broadcast_i32x8 (__mmask16 __M, __m256i __A) |
363 | { |
364 | return (__m512i) __builtin_ia32_broadcasti32x8_512_mask ((__v8si) |
365 | __A, |
366 | (__v16si) |
367 | _mm512_setzero_si512 (), |
368 | __M); |
369 | } |
370 | |
371 | extern __inline __m512i |
372 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
373 | _mm512_mullo_epi64 (__m512i __A, __m512i __B) |
374 | { |
375 | return (__m512i) ((__v8du) __A * (__v8du) __B); |
376 | } |
377 | |
378 | extern __inline __m512i |
379 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
380 | _mm512_mask_mullo_epi64 (__m512i __W, __mmask8 __U, __m512i __A, |
381 | __m512i __B) |
382 | { |
383 | return (__m512i) __builtin_ia32_pmullq512_mask ((__v8di) __A, |
384 | (__v8di) __B, |
385 | (__v8di) __W, |
386 | (__mmask8) __U); |
387 | } |
388 | |
389 | extern __inline __m512i |
390 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
391 | _mm512_maskz_mullo_epi64 (__mmask8 __U, __m512i __A, __m512i __B) |
392 | { |
393 | return (__m512i) __builtin_ia32_pmullq512_mask ((__v8di) __A, |
394 | (__v8di) __B, |
395 | (__v8di) |
396 | _mm512_setzero_si512 (), |
397 | (__mmask8) __U); |
398 | } |
399 | |
400 | extern __inline __m512d |
401 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
402 | _mm512_xor_pd (__m512d __A, __m512d __B) |
403 | { |
404 | return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, |
405 | (__v8df) __B, |
406 | (__v8df) |
407 | _mm512_setzero_pd (), |
408 | (__mmask8) -1); |
409 | } |
410 | |
411 | extern __inline __m512d |
412 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
413 | _mm512_mask_xor_pd (__m512d __W, __mmask8 __U, __m512d __A, |
414 | __m512d __B) |
415 | { |
416 | return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, |
417 | (__v8df) __B, |
418 | (__v8df) __W, |
419 | (__mmask8) __U); |
420 | } |
421 | |
422 | extern __inline __m512d |
423 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
424 | _mm512_maskz_xor_pd (__mmask8 __U, __m512d __A, __m512d __B) |
425 | { |
426 | return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, |
427 | (__v8df) __B, |
428 | (__v8df) |
429 | _mm512_setzero_pd (), |
430 | (__mmask8) __U); |
431 | } |
432 | |
433 | extern __inline __m512 |
434 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
435 | _mm512_xor_ps (__m512 __A, __m512 __B) |
436 | { |
437 | return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, |
438 | (__v16sf) __B, |
439 | (__v16sf) |
440 | _mm512_setzero_ps (), |
441 | (__mmask16) -1); |
442 | } |
443 | |
444 | extern __inline __m512 |
445 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
446 | _mm512_mask_xor_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) |
447 | { |
448 | return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, |
449 | (__v16sf) __B, |
450 | (__v16sf) __W, |
451 | (__mmask16) __U); |
452 | } |
453 | |
454 | extern __inline __m512 |
455 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
456 | _mm512_maskz_xor_ps (__mmask16 __U, __m512 __A, __m512 __B) |
457 | { |
458 | return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, |
459 | (__v16sf) __B, |
460 | (__v16sf) |
461 | _mm512_setzero_ps (), |
462 | (__mmask16) __U); |
463 | } |
464 | |
465 | extern __inline __m512d |
466 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
467 | _mm512_or_pd (__m512d __A, __m512d __B) |
468 | { |
469 | return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, |
470 | (__v8df) __B, |
471 | (__v8df) |
472 | _mm512_setzero_pd (), |
473 | (__mmask8) -1); |
474 | } |
475 | |
476 | extern __inline __m512d |
477 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
478 | _mm512_mask_or_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) |
479 | { |
480 | return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, |
481 | (__v8df) __B, |
482 | (__v8df) __W, |
483 | (__mmask8) __U); |
484 | } |
485 | |
486 | extern __inline __m512d |
487 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
488 | _mm512_maskz_or_pd (__mmask8 __U, __m512d __A, __m512d __B) |
489 | { |
490 | return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, |
491 | (__v8df) __B, |
492 | (__v8df) |
493 | _mm512_setzero_pd (), |
494 | (__mmask8) __U); |
495 | } |
496 | |
497 | extern __inline __m512 |
498 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
499 | _mm512_or_ps (__m512 __A, __m512 __B) |
500 | { |
501 | return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, |
502 | (__v16sf) __B, |
503 | (__v16sf) |
504 | _mm512_setzero_ps (), |
505 | (__mmask16) -1); |
506 | } |
507 | |
508 | extern __inline __m512 |
509 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
510 | _mm512_mask_or_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) |
511 | { |
512 | return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, |
513 | (__v16sf) __B, |
514 | (__v16sf) __W, |
515 | (__mmask16) __U); |
516 | } |
517 | |
518 | extern __inline __m512 |
519 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
520 | _mm512_maskz_or_ps (__mmask16 __U, __m512 __A, __m512 __B) |
521 | { |
522 | return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, |
523 | (__v16sf) __B, |
524 | (__v16sf) |
525 | _mm512_setzero_ps (), |
526 | (__mmask16) __U); |
527 | } |
528 | |
529 | extern __inline __m512d |
530 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
531 | _mm512_and_pd (__m512d __A, __m512d __B) |
532 | { |
533 | return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, |
534 | (__v8df) __B, |
535 | (__v8df) |
536 | _mm512_setzero_pd (), |
537 | (__mmask8) -1); |
538 | } |
539 | |
540 | extern __inline __m512d |
541 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
542 | _mm512_mask_and_pd (__m512d __W, __mmask8 __U, __m512d __A, |
543 | __m512d __B) |
544 | { |
545 | return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, |
546 | (__v8df) __B, |
547 | (__v8df) __W, |
548 | (__mmask8) __U); |
549 | } |
550 | |
551 | extern __inline __m512d |
552 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
553 | _mm512_maskz_and_pd (__mmask8 __U, __m512d __A, __m512d __B) |
554 | { |
555 | return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, |
556 | (__v8df) __B, |
557 | (__v8df) |
558 | _mm512_setzero_pd (), |
559 | (__mmask8) __U); |
560 | } |
561 | |
562 | extern __inline __m512 |
563 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
564 | _mm512_and_ps (__m512 __A, __m512 __B) |
565 | { |
566 | return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, |
567 | (__v16sf) __B, |
568 | (__v16sf) |
569 | _mm512_setzero_ps (), |
570 | (__mmask16) -1); |
571 | } |
572 | |
573 | extern __inline __m512 |
574 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
575 | _mm512_mask_and_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) |
576 | { |
577 | return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, |
578 | (__v16sf) __B, |
579 | (__v16sf) __W, |
580 | (__mmask16) __U); |
581 | } |
582 | |
583 | extern __inline __m512 |
584 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
585 | _mm512_maskz_and_ps (__mmask16 __U, __m512 __A, __m512 __B) |
586 | { |
587 | return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, |
588 | (__v16sf) __B, |
589 | (__v16sf) |
590 | _mm512_setzero_ps (), |
591 | (__mmask16) __U); |
592 | } |
593 | |
594 | extern __inline __m512d |
595 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
596 | _mm512_andnot_pd (__m512d __A, __m512d __B) |
597 | { |
598 | return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, |
599 | (__v8df) __B, |
600 | (__v8df) |
601 | _mm512_setzero_pd (), |
602 | (__mmask8) -1); |
603 | } |
604 | |
605 | extern __inline __m512d |
606 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
607 | _mm512_mask_andnot_pd (__m512d __W, __mmask8 __U, __m512d __A, |
608 | __m512d __B) |
609 | { |
610 | return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, |
611 | (__v8df) __B, |
612 | (__v8df) __W, |
613 | (__mmask8) __U); |
614 | } |
615 | |
616 | extern __inline __m512d |
617 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
618 | _mm512_maskz_andnot_pd (__mmask8 __U, __m512d __A, __m512d __B) |
619 | { |
620 | return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, |
621 | (__v8df) __B, |
622 | (__v8df) |
623 | _mm512_setzero_pd (), |
624 | (__mmask8) __U); |
625 | } |
626 | |
627 | extern __inline __m512 |
628 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
629 | _mm512_andnot_ps (__m512 __A, __m512 __B) |
630 | { |
631 | return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, |
632 | (__v16sf) __B, |
633 | (__v16sf) |
634 | _mm512_setzero_ps (), |
635 | (__mmask16) -1); |
636 | } |
637 | |
638 | extern __inline __m512 |
639 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
640 | _mm512_mask_andnot_ps (__m512 __W, __mmask16 __U, __m512 __A, |
641 | __m512 __B) |
642 | { |
643 | return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, |
644 | (__v16sf) __B, |
645 | (__v16sf) __W, |
646 | (__mmask16) __U); |
647 | } |
648 | |
649 | extern __inline __m512 |
650 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
651 | _mm512_maskz_andnot_ps (__mmask16 __U, __m512 __A, __m512 __B) |
652 | { |
653 | return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, |
654 | (__v16sf) __B, |
655 | (__v16sf) |
656 | _mm512_setzero_ps (), |
657 | (__mmask16) __U); |
658 | } |
659 | |
660 | extern __inline __mmask16 |
661 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
662 | _mm512_movepi32_mask (__m512i __A) |
663 | { |
664 | return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A); |
665 | } |
666 | |
667 | extern __inline __mmask8 |
668 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
669 | _mm512_movepi64_mask (__m512i __A) |
670 | { |
671 | return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A); |
672 | } |
673 | |
674 | extern __inline __m512i |
675 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
676 | _mm512_movm_epi32 (__mmask16 __A) |
677 | { |
678 | return (__m512i) __builtin_ia32_cvtmask2d512 (__A); |
679 | } |
680 | |
681 | extern __inline __m512i |
682 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
683 | _mm512_movm_epi64 (__mmask8 __A) |
684 | { |
685 | return (__m512i) __builtin_ia32_cvtmask2q512 (__A); |
686 | } |
687 | |
688 | extern __inline __m512i |
689 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
690 | _mm512_cvttpd_epi64 (__m512d __A) |
691 | { |
692 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
693 | (__v8di) |
694 | _mm512_setzero_si512 (), |
695 | (__mmask8) -1, |
696 | _MM_FROUND_CUR_DIRECTION); |
697 | } |
698 | |
699 | extern __inline __m512i |
700 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
701 | _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) |
702 | { |
703 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
704 | (__v8di) __W, |
705 | (__mmask8) __U, |
706 | _MM_FROUND_CUR_DIRECTION); |
707 | } |
708 | |
709 | extern __inline __m512i |
710 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
711 | _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) |
712 | { |
713 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
714 | (__v8di) |
715 | _mm512_setzero_si512 (), |
716 | (__mmask8) __U, |
717 | _MM_FROUND_CUR_DIRECTION); |
718 | } |
719 | |
720 | extern __inline __m512i |
721 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
722 | _mm512_cvttpd_epu64 (__m512d __A) |
723 | { |
724 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
725 | (__v8di) |
726 | _mm512_setzero_si512 (), |
727 | (__mmask8) -1, |
728 | _MM_FROUND_CUR_DIRECTION); |
729 | } |
730 | |
731 | extern __inline __m512i |
732 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
733 | _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) |
734 | { |
735 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
736 | (__v8di) __W, |
737 | (__mmask8) __U, |
738 | _MM_FROUND_CUR_DIRECTION); |
739 | } |
740 | |
741 | extern __inline __m512i |
742 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
743 | _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) |
744 | { |
745 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
746 | (__v8di) |
747 | _mm512_setzero_si512 (), |
748 | (__mmask8) __U, |
749 | _MM_FROUND_CUR_DIRECTION); |
750 | } |
751 | |
752 | extern __inline __m512i |
753 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
754 | _mm512_cvttps_epi64 (__m256 __A) |
755 | { |
756 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
757 | (__v8di) |
758 | _mm512_setzero_si512 (), |
759 | (__mmask8) -1, |
760 | _MM_FROUND_CUR_DIRECTION); |
761 | } |
762 | |
763 | extern __inline __m512i |
764 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
765 | _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) |
766 | { |
767 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
768 | (__v8di) __W, |
769 | (__mmask8) __U, |
770 | _MM_FROUND_CUR_DIRECTION); |
771 | } |
772 | |
773 | extern __inline __m512i |
774 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
775 | _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) |
776 | { |
777 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
778 | (__v8di) |
779 | _mm512_setzero_si512 (), |
780 | (__mmask8) __U, |
781 | _MM_FROUND_CUR_DIRECTION); |
782 | } |
783 | |
784 | extern __inline __m512i |
785 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
786 | _mm512_cvttps_epu64 (__m256 __A) |
787 | { |
788 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
789 | (__v8di) |
790 | _mm512_setzero_si512 (), |
791 | (__mmask8) -1, |
792 | _MM_FROUND_CUR_DIRECTION); |
793 | } |
794 | |
795 | extern __inline __m512i |
796 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
797 | _mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) |
798 | { |
799 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
800 | (__v8di) __W, |
801 | (__mmask8) __U, |
802 | _MM_FROUND_CUR_DIRECTION); |
803 | } |
804 | |
805 | extern __inline __m512i |
806 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
807 | _mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) |
808 | { |
809 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
810 | (__v8di) |
811 | _mm512_setzero_si512 (), |
812 | (__mmask8) __U, |
813 | _MM_FROUND_CUR_DIRECTION); |
814 | } |
815 | |
816 | extern __inline __m512i |
817 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
818 | _mm512_cvtpd_epi64 (__m512d __A) |
819 | { |
820 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
821 | (__v8di) |
822 | _mm512_setzero_si512 (), |
823 | (__mmask8) -1, |
824 | _MM_FROUND_CUR_DIRECTION); |
825 | } |
826 | |
827 | extern __inline __m512i |
828 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
829 | _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) |
830 | { |
831 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
832 | (__v8di) __W, |
833 | (__mmask8) __U, |
834 | _MM_FROUND_CUR_DIRECTION); |
835 | } |
836 | |
837 | extern __inline __m512i |
838 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
839 | _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) |
840 | { |
841 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
842 | (__v8di) |
843 | _mm512_setzero_si512 (), |
844 | (__mmask8) __U, |
845 | _MM_FROUND_CUR_DIRECTION); |
846 | } |
847 | |
848 | extern __inline __m512i |
849 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
850 | _mm512_cvtpd_epu64 (__m512d __A) |
851 | { |
852 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
853 | (__v8di) |
854 | _mm512_setzero_si512 (), |
855 | (__mmask8) -1, |
856 | _MM_FROUND_CUR_DIRECTION); |
857 | } |
858 | |
859 | extern __inline __m512i |
860 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
861 | _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) |
862 | { |
863 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
864 | (__v8di) __W, |
865 | (__mmask8) __U, |
866 | _MM_FROUND_CUR_DIRECTION); |
867 | } |
868 | |
869 | extern __inline __m512i |
870 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
871 | _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) |
872 | { |
873 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
874 | (__v8di) |
875 | _mm512_setzero_si512 (), |
876 | (__mmask8) __U, |
877 | _MM_FROUND_CUR_DIRECTION); |
878 | } |
879 | |
880 | extern __inline __m512i |
881 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
882 | _mm512_cvtps_epi64 (__m256 __A) |
883 | { |
884 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
885 | (__v8di) |
886 | _mm512_setzero_si512 (), |
887 | (__mmask8) -1, |
888 | _MM_FROUND_CUR_DIRECTION); |
889 | } |
890 | |
891 | extern __inline __m512i |
892 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
893 | _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) |
894 | { |
895 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
896 | (__v8di) __W, |
897 | (__mmask8) __U, |
898 | _MM_FROUND_CUR_DIRECTION); |
899 | } |
900 | |
901 | extern __inline __m512i |
902 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
903 | _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) |
904 | { |
905 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
906 | (__v8di) |
907 | _mm512_setzero_si512 (), |
908 | (__mmask8) __U, |
909 | _MM_FROUND_CUR_DIRECTION); |
910 | } |
911 | |
912 | extern __inline __m512i |
913 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
914 | _mm512_cvtps_epu64 (__m256 __A) |
915 | { |
916 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
917 | (__v8di) |
918 | _mm512_setzero_si512 (), |
919 | (__mmask8) -1, |
920 | _MM_FROUND_CUR_DIRECTION); |
921 | } |
922 | |
923 | extern __inline __m512i |
924 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
925 | _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) |
926 | { |
927 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
928 | (__v8di) __W, |
929 | (__mmask8) __U, |
930 | _MM_FROUND_CUR_DIRECTION); |
931 | } |
932 | |
933 | extern __inline __m512i |
934 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
935 | _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) |
936 | { |
937 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
938 | (__v8di) |
939 | _mm512_setzero_si512 (), |
940 | (__mmask8) __U, |
941 | _MM_FROUND_CUR_DIRECTION); |
942 | } |
943 | |
944 | extern __inline __m256 |
945 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
946 | _mm512_cvtepi64_ps (__m512i __A) |
947 | { |
948 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
949 | (__v8sf) |
950 | _mm256_setzero_ps (), |
951 | (__mmask8) -1, |
952 | _MM_FROUND_CUR_DIRECTION); |
953 | } |
954 | |
955 | extern __inline __m256 |
956 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
957 | _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) |
958 | { |
959 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
960 | (__v8sf) __W, |
961 | (__mmask8) __U, |
962 | _MM_FROUND_CUR_DIRECTION); |
963 | } |
964 | |
965 | extern __inline __m256 |
966 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
967 | _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) |
968 | { |
969 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
970 | (__v8sf) |
971 | _mm256_setzero_ps (), |
972 | (__mmask8) __U, |
973 | _MM_FROUND_CUR_DIRECTION); |
974 | } |
975 | |
976 | extern __inline __m256 |
977 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
978 | _mm512_cvtepu64_ps (__m512i __A) |
979 | { |
980 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
981 | (__v8sf) |
982 | _mm256_setzero_ps (), |
983 | (__mmask8) -1, |
984 | _MM_FROUND_CUR_DIRECTION); |
985 | } |
986 | |
987 | extern __inline __m256 |
988 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
989 | _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) |
990 | { |
991 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
992 | (__v8sf) __W, |
993 | (__mmask8) __U, |
994 | _MM_FROUND_CUR_DIRECTION); |
995 | } |
996 | |
997 | extern __inline __m256 |
998 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
999 | _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) |
1000 | { |
1001 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
1002 | (__v8sf) |
1003 | _mm256_setzero_ps (), |
1004 | (__mmask8) __U, |
1005 | _MM_FROUND_CUR_DIRECTION); |
1006 | } |
1007 | |
1008 | extern __inline __m512d |
1009 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1010 | _mm512_cvtepi64_pd (__m512i __A) |
1011 | { |
1012 | return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, |
1013 | (__v8df) |
1014 | _mm512_setzero_pd (), |
1015 | (__mmask8) -1, |
1016 | _MM_FROUND_CUR_DIRECTION); |
1017 | } |
1018 | |
1019 | extern __inline __m512d |
1020 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1021 | _mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) |
1022 | { |
1023 | return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, |
1024 | (__v8df) __W, |
1025 | (__mmask8) __U, |
1026 | _MM_FROUND_CUR_DIRECTION); |
1027 | } |
1028 | |
1029 | extern __inline __m512d |
1030 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1031 | _mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) |
1032 | { |
1033 | return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, |
1034 | (__v8df) |
1035 | _mm512_setzero_pd (), |
1036 | (__mmask8) __U, |
1037 | _MM_FROUND_CUR_DIRECTION); |
1038 | } |
1039 | |
1040 | extern __inline __m512d |
1041 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1042 | _mm512_cvtepu64_pd (__m512i __A) |
1043 | { |
1044 | return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, |
1045 | (__v8df) |
1046 | _mm512_setzero_pd (), |
1047 | (__mmask8) -1, |
1048 | _MM_FROUND_CUR_DIRECTION); |
1049 | } |
1050 | |
1051 | extern __inline __m512d |
1052 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1053 | _mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) |
1054 | { |
1055 | return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, |
1056 | (__v8df) __W, |
1057 | (__mmask8) __U, |
1058 | _MM_FROUND_CUR_DIRECTION); |
1059 | } |
1060 | |
1061 | extern __inline __m512d |
1062 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1063 | _mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) |
1064 | { |
1065 | return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, |
1066 | (__v8df) |
1067 | _mm512_setzero_pd (), |
1068 | (__mmask8) __U, |
1069 | _MM_FROUND_CUR_DIRECTION); |
1070 | } |
1071 | |
1072 | #ifdef __OPTIMIZE__ |
1073 | extern __inline __mmask8 |
1074 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1075 | _kshiftli_mask8 (__mmask8 __A, unsigned int __B) |
1076 | { |
1077 | return (__mmask8) __builtin_ia32_kshiftliqi ((__mmask8) __A, (__mmask8) __B); |
1078 | } |
1079 | |
1080 | extern __inline __mmask8 |
1081 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1082 | _kshiftri_mask8 (__mmask8 __A, unsigned int __B) |
1083 | { |
1084 | return (__mmask8) __builtin_ia32_kshiftriqi ((__mmask8) __A, (__mmask8) __B); |
1085 | } |
1086 | |
1087 | extern __inline __m512d |
1088 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1089 | _mm512_range_pd (__m512d __A, __m512d __B, int __C) |
1090 | { |
1091 | return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, |
1092 | (__v8df) __B, __C, |
1093 | (__v8df) |
1094 | _mm512_setzero_pd (), |
1095 | (__mmask8) -1, |
1096 | _MM_FROUND_CUR_DIRECTION); |
1097 | } |
1098 | |
1099 | extern __inline __m512d |
1100 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1101 | _mm512_mask_range_pd (__m512d __W, __mmask8 __U, |
1102 | __m512d __A, __m512d __B, int __C) |
1103 | { |
1104 | return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, |
1105 | (__v8df) __B, __C, |
1106 | (__v8df) __W, |
1107 | (__mmask8) __U, |
1108 | _MM_FROUND_CUR_DIRECTION); |
1109 | } |
1110 | |
1111 | extern __inline __m512d |
1112 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1113 | _mm512_maskz_range_pd (__mmask8 __U, __m512d __A, __m512d __B, int __C) |
1114 | { |
1115 | return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, |
1116 | (__v8df) __B, __C, |
1117 | (__v8df) |
1118 | _mm512_setzero_pd (), |
1119 | (__mmask8) __U, |
1120 | _MM_FROUND_CUR_DIRECTION); |
1121 | } |
1122 | |
1123 | extern __inline __m512 |
1124 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1125 | _mm512_range_ps (__m512 __A, __m512 __B, int __C) |
1126 | { |
1127 | return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, |
1128 | (__v16sf) __B, __C, |
1129 | (__v16sf) |
1130 | _mm512_setzero_ps (), |
1131 | (__mmask16) -1, |
1132 | _MM_FROUND_CUR_DIRECTION); |
1133 | } |
1134 | |
1135 | extern __inline __m512 |
1136 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1137 | _mm512_mask_range_ps (__m512 __W, __mmask16 __U, |
1138 | __m512 __A, __m512 __B, int __C) |
1139 | { |
1140 | return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, |
1141 | (__v16sf) __B, __C, |
1142 | (__v16sf) __W, |
1143 | (__mmask16) __U, |
1144 | _MM_FROUND_CUR_DIRECTION); |
1145 | } |
1146 | |
1147 | extern __inline __m512 |
1148 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1149 | _mm512_maskz_range_ps (__mmask16 __U, __m512 __A, __m512 __B, int __C) |
1150 | { |
1151 | return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, |
1152 | (__v16sf) __B, __C, |
1153 | (__v16sf) |
1154 | _mm512_setzero_ps (), |
1155 | (__mmask16) __U, |
1156 | _MM_FROUND_CUR_DIRECTION); |
1157 | } |
1158 | |
1159 | extern __inline __m128d |
1160 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1161 | _mm_reduce_sd (__m128d __A, __m128d __B, int __C) |
1162 | { |
1163 | return (__m128d) __builtin_ia32_reducesd_mask ((__v2df) __A, |
1164 | (__v2df) __B, __C, |
1165 | (__v2df) _mm_setzero_pd (), |
1166 | (__mmask8) -1); |
1167 | } |
1168 | |
1169 | extern __inline __m128d |
1170 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1171 | _mm_mask_reduce_sd (__m128d __W, __mmask8 __U, __m128d __A, |
1172 | __m128d __B, int __C) |
1173 | { |
1174 | return (__m128d) __builtin_ia32_reducesd_mask ((__v2df) __A, |
1175 | (__v2df) __B, __C, |
1176 | (__v2df) __W, |
1177 | (__mmask8) __U); |
1178 | } |
1179 | |
1180 | extern __inline __m128d |
1181 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1182 | _mm_maskz_reduce_sd (__mmask8 __U, __m128d __A, __m128d __B, int __C) |
1183 | { |
1184 | return (__m128d) __builtin_ia32_reducesd_mask ((__v2df) __A, |
1185 | (__v2df) __B, __C, |
1186 | (__v2df) _mm_setzero_pd (), |
1187 | (__mmask8) __U); |
1188 | } |
1189 | |
1190 | extern __inline __m128 |
1191 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1192 | _mm_reduce_ss (__m128 __A, __m128 __B, int __C) |
1193 | { |
1194 | return (__m128) __builtin_ia32_reducess_mask ((__v4sf) __A, |
1195 | (__v4sf) __B, __C, |
1196 | (__v4sf) _mm_setzero_ps (), |
1197 | (__mmask8) -1); |
1198 | } |
1199 | |
1200 | |
1201 | extern __inline __m128 |
1202 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1203 | _mm_mask_reduce_ss (__m128 __W, __mmask8 __U, __m128 __A, |
1204 | __m128 __B, int __C) |
1205 | { |
1206 | return (__m128) __builtin_ia32_reducess_mask ((__v4sf) __A, |
1207 | (__v4sf) __B, __C, |
1208 | (__v4sf) __W, |
1209 | (__mmask8) __U); |
1210 | } |
1211 | |
1212 | extern __inline __m128 |
1213 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1214 | _mm_maskz_reduce_ss (__mmask8 __U, __m128 __A, __m128 __B, int __C) |
1215 | { |
1216 | return (__m128) __builtin_ia32_reducess_mask ((__v4sf) __A, |
1217 | (__v4sf) __B, __C, |
1218 | (__v4sf) _mm_setzero_ps (), |
1219 | (__mmask8) __U); |
1220 | } |
1221 | |
1222 | extern __inline __m128d |
1223 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1224 | _mm_range_sd (__m128d __A, __m128d __B, int __C) |
1225 | { |
1226 | return (__m128d) __builtin_ia32_rangesd128_round ((__v2df) __A, |
1227 | (__v2df) __B, __C, |
1228 | _MM_FROUND_CUR_DIRECTION); |
1229 | } |
1230 | |
1231 | |
1232 | extern __inline __m128 |
1233 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1234 | _mm_range_ss (__m128 __A, __m128 __B, int __C) |
1235 | { |
1236 | return (__m128) __builtin_ia32_rangess128_round ((__v4sf) __A, |
1237 | (__v4sf) __B, __C, |
1238 | _MM_FROUND_CUR_DIRECTION); |
1239 | } |
1240 | |
1241 | extern __inline __m128d |
1242 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1243 | _mm_range_round_sd (__m128d __A, __m128d __B, int __C, const int __R) |
1244 | { |
1245 | return (__m128d) __builtin_ia32_rangesd128_round ((__v2df) __A, |
1246 | (__v2df) __B, __C, |
1247 | __R); |
1248 | } |
1249 | |
1250 | extern __inline __m128 |
1251 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1252 | _mm_range_round_ss (__m128 __A, __m128 __B, int __C, const int __R) |
1253 | { |
1254 | return (__m128) __builtin_ia32_rangess128_round ((__v4sf) __A, |
1255 | (__v4sf) __B, __C, |
1256 | __R); |
1257 | } |
1258 | |
1259 | extern __inline __mmask8 |
1260 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1261 | _mm_fpclass_ss_mask (__m128 __A, const int __imm) |
1262 | { |
1263 | return (__mmask8) __builtin_ia32_fpclassss ((__v4sf) __A, __imm); |
1264 | } |
1265 | |
1266 | extern __inline __mmask8 |
1267 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1268 | _mm_fpclass_sd_mask (__m128d __A, const int __imm) |
1269 | { |
1270 | return (__mmask8) __builtin_ia32_fpclasssd ((__v2df) __A, __imm); |
1271 | } |
1272 | |
1273 | extern __inline __m512i |
1274 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1275 | _mm512_cvtt_roundpd_epi64 (__m512d __A, const int __R) |
1276 | { |
1277 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
1278 | (__v8di) |
1279 | _mm512_setzero_si512 (), |
1280 | (__mmask8) -1, |
1281 | __R); |
1282 | } |
1283 | |
1284 | extern __inline __m512i |
1285 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1286 | _mm512_mask_cvtt_roundpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A, |
1287 | const int __R) |
1288 | { |
1289 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
1290 | (__v8di) __W, |
1291 | (__mmask8) __U, |
1292 | __R); |
1293 | } |
1294 | |
1295 | extern __inline __m512i |
1296 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1297 | _mm512_maskz_cvtt_roundpd_epi64 (__mmask8 __U, __m512d __A, |
1298 | const int __R) |
1299 | { |
1300 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
1301 | (__v8di) |
1302 | _mm512_setzero_si512 (), |
1303 | (__mmask8) __U, |
1304 | __R); |
1305 | } |
1306 | |
1307 | extern __inline __m512i |
1308 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1309 | _mm512_cvtt_roundpd_epu64 (__m512d __A, const int __R) |
1310 | { |
1311 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
1312 | (__v8di) |
1313 | _mm512_setzero_si512 (), |
1314 | (__mmask8) -1, |
1315 | __R); |
1316 | } |
1317 | |
1318 | extern __inline __m512i |
1319 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1320 | _mm512_mask_cvtt_roundpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A, |
1321 | const int __R) |
1322 | { |
1323 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
1324 | (__v8di) __W, |
1325 | (__mmask8) __U, |
1326 | __R); |
1327 | } |
1328 | |
1329 | extern __inline __m512i |
1330 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1331 | _mm512_maskz_cvtt_roundpd_epu64 (__mmask8 __U, __m512d __A, |
1332 | const int __R) |
1333 | { |
1334 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
1335 | (__v8di) |
1336 | _mm512_setzero_si512 (), |
1337 | (__mmask8) __U, |
1338 | __R); |
1339 | } |
1340 | |
1341 | extern __inline __m512i |
1342 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1343 | _mm512_cvtt_roundps_epi64 (__m256 __A, const int __R) |
1344 | { |
1345 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
1346 | (__v8di) |
1347 | _mm512_setzero_si512 (), |
1348 | (__mmask8) -1, |
1349 | __R); |
1350 | } |
1351 | |
1352 | extern __inline __m512i |
1353 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1354 | _mm512_mask_cvtt_roundps_epi64 (__m512i __W, __mmask8 __U, __m256 __A, |
1355 | const int __R) |
1356 | { |
1357 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
1358 | (__v8di) __W, |
1359 | (__mmask8) __U, |
1360 | __R); |
1361 | } |
1362 | |
1363 | extern __inline __m512i |
1364 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1365 | _mm512_maskz_cvtt_roundps_epi64 (__mmask8 __U, __m256 __A, |
1366 | const int __R) |
1367 | { |
1368 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
1369 | (__v8di) |
1370 | _mm512_setzero_si512 (), |
1371 | (__mmask8) __U, |
1372 | __R); |
1373 | } |
1374 | |
1375 | extern __inline __m512i |
1376 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1377 | _mm512_cvtt_roundps_epu64 (__m256 __A, const int __R) |
1378 | { |
1379 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
1380 | (__v8di) |
1381 | _mm512_setzero_si512 (), |
1382 | (__mmask8) -1, |
1383 | __R); |
1384 | } |
1385 | |
1386 | extern __inline __m512i |
1387 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1388 | _mm512_mask_cvtt_roundps_epu64 (__m512i __W, __mmask8 __U, __m256 __A, |
1389 | const int __R) |
1390 | { |
1391 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
1392 | (__v8di) __W, |
1393 | (__mmask8) __U, |
1394 | __R); |
1395 | } |
1396 | |
1397 | extern __inline __m512i |
1398 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1399 | _mm512_maskz_cvtt_roundps_epu64 (__mmask8 __U, __m256 __A, |
1400 | const int __R) |
1401 | { |
1402 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
1403 | (__v8di) |
1404 | _mm512_setzero_si512 (), |
1405 | (__mmask8) __U, |
1406 | __R); |
1407 | } |
1408 | |
1409 | extern __inline __m512i |
1410 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1411 | _mm512_cvt_roundpd_epi64 (__m512d __A, const int __R) |
1412 | { |
1413 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
1414 | (__v8di) |
1415 | _mm512_setzero_si512 (), |
1416 | (__mmask8) -1, |
1417 | __R); |
1418 | } |
1419 | |
1420 | extern __inline __m512i |
1421 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1422 | _mm512_mask_cvt_roundpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A, |
1423 | const int __R) |
1424 | { |
1425 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
1426 | (__v8di) __W, |
1427 | (__mmask8) __U, |
1428 | __R); |
1429 | } |
1430 | |
1431 | extern __inline __m512i |
1432 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1433 | _mm512_maskz_cvt_roundpd_epi64 (__mmask8 __U, __m512d __A, |
1434 | const int __R) |
1435 | { |
1436 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
1437 | (__v8di) |
1438 | _mm512_setzero_si512 (), |
1439 | (__mmask8) __U, |
1440 | __R); |
1441 | } |
1442 | |
1443 | extern __inline __m512i |
1444 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1445 | _mm512_cvt_roundpd_epu64 (__m512d __A, const int __R) |
1446 | { |
1447 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
1448 | (__v8di) |
1449 | _mm512_setzero_si512 (), |
1450 | (__mmask8) -1, |
1451 | __R); |
1452 | } |
1453 | |
1454 | extern __inline __m512i |
1455 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1456 | _mm512_mask_cvt_roundpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A, |
1457 | const int __R) |
1458 | { |
1459 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
1460 | (__v8di) __W, |
1461 | (__mmask8) __U, |
1462 | __R); |
1463 | } |
1464 | |
1465 | extern __inline __m512i |
1466 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1467 | _mm512_maskz_cvt_roundpd_epu64 (__mmask8 __U, __m512d __A, |
1468 | const int __R) |
1469 | { |
1470 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
1471 | (__v8di) |
1472 | _mm512_setzero_si512 (), |
1473 | (__mmask8) __U, |
1474 | __R); |
1475 | } |
1476 | |
1477 | extern __inline __m512i |
1478 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1479 | _mm512_cvt_roundps_epi64 (__m256 __A, const int __R) |
1480 | { |
1481 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
1482 | (__v8di) |
1483 | _mm512_setzero_si512 (), |
1484 | (__mmask8) -1, |
1485 | __R); |
1486 | } |
1487 | |
1488 | extern __inline __m512i |
1489 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1490 | _mm512_mask_cvt_roundps_epi64 (__m512i __W, __mmask8 __U, __m256 __A, |
1491 | const int __R) |
1492 | { |
1493 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
1494 | (__v8di) __W, |
1495 | (__mmask8) __U, |
1496 | __R); |
1497 | } |
1498 | |
1499 | extern __inline __m512i |
1500 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1501 | _mm512_maskz_cvt_roundps_epi64 (__mmask8 __U, __m256 __A, |
1502 | const int __R) |
1503 | { |
1504 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
1505 | (__v8di) |
1506 | _mm512_setzero_si512 (), |
1507 | (__mmask8) __U, |
1508 | __R); |
1509 | } |
1510 | |
1511 | extern __inline __m512i |
1512 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1513 | _mm512_cvt_roundps_epu64 (__m256 __A, const int __R) |
1514 | { |
1515 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
1516 | (__v8di) |
1517 | _mm512_setzero_si512 (), |
1518 | (__mmask8) -1, |
1519 | __R); |
1520 | } |
1521 | |
1522 | extern __inline __m512i |
1523 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1524 | _mm512_mask_cvt_roundps_epu64 (__m512i __W, __mmask8 __U, __m256 __A, |
1525 | const int __R) |
1526 | { |
1527 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
1528 | (__v8di) __W, |
1529 | (__mmask8) __U, |
1530 | __R); |
1531 | } |
1532 | |
1533 | extern __inline __m512i |
1534 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1535 | _mm512_maskz_cvt_roundps_epu64 (__mmask8 __U, __m256 __A, |
1536 | const int __R) |
1537 | { |
1538 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
1539 | (__v8di) |
1540 | _mm512_setzero_si512 (), |
1541 | (__mmask8) __U, |
1542 | __R); |
1543 | } |
1544 | |
1545 | extern __inline __m256 |
1546 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1547 | _mm512_cvt_roundepi64_ps (__m512i __A, const int __R) |
1548 | { |
1549 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
1550 | (__v8sf) |
1551 | _mm256_setzero_ps (), |
1552 | (__mmask8) -1, |
1553 | __R); |
1554 | } |
1555 | |
1556 | extern __inline __m256 |
1557 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1558 | _mm512_mask_cvt_roundepi64_ps (__m256 __W, __mmask8 __U, __m512i __A, |
1559 | const int __R) |
1560 | { |
1561 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
1562 | (__v8sf) __W, |
1563 | (__mmask8) __U, |
1564 | __R); |
1565 | } |
1566 | |
1567 | extern __inline __m256 |
1568 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1569 | _mm512_maskz_cvt_roundepi64_ps (__mmask8 __U, __m512i __A, |
1570 | const int __R) |
1571 | { |
1572 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
1573 | (__v8sf) |
1574 | _mm256_setzero_ps (), |
1575 | (__mmask8) __U, |
1576 | __R); |
1577 | } |
1578 | |
1579 | extern __inline __m256 |
1580 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1581 | _mm512_cvt_roundepu64_ps (__m512i __A, const int __R) |
1582 | { |
1583 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
1584 | (__v8sf) |
1585 | _mm256_setzero_ps (), |
1586 | (__mmask8) -1, |
1587 | __R); |
1588 | } |
1589 | |
1590 | extern __inline __m256 |
1591 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1592 | _mm512_mask_cvt_roundepu64_ps (__m256 __W, __mmask8 __U, __m512i __A, |
1593 | const int __R) |
1594 | { |
1595 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
1596 | (__v8sf) __W, |
1597 | (__mmask8) __U, |
1598 | __R); |
1599 | } |
1600 | |
1601 | extern __inline __m256 |
1602 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1603 | _mm512_maskz_cvt_roundepu64_ps (__mmask8 __U, __m512i __A, |
1604 | const int __R) |
1605 | { |
1606 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
1607 | (__v8sf) |
1608 | _mm256_setzero_ps (), |
1609 | (__mmask8) __U, |
1610 | __R); |
1611 | } |
1612 | |
1613 | extern __inline __m512d |
1614 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1615 | _mm512_cvt_roundepi64_pd (__m512i __A, const int __R) |
1616 | { |
1617 | return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, |
1618 | (__v8df) |
1619 | _mm512_setzero_pd (), |
1620 | (__mmask8) -1, |
1621 | __R); |
1622 | } |
1623 | |
1624 | extern __inline __m512d |
1625 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1626 | _mm512_mask_cvt_roundepi64_pd (__m512d __W, __mmask8 __U, __m512i __A, |
1627 | const int __R) |
1628 | { |
1629 | return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, |
1630 | (__v8df) __W, |
1631 | (__mmask8) __U, |
1632 | __R); |
1633 | } |
1634 | |
1635 | extern __inline __m512d |
1636 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1637 | _mm512_maskz_cvt_roundepi64_pd (__mmask8 __U, __m512i __A, |
1638 | const int __R) |
1639 | { |
1640 | return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, |
1641 | (__v8df) |
1642 | _mm512_setzero_pd (), |
1643 | (__mmask8) __U, |
1644 | __R); |
1645 | } |
1646 | |
1647 | extern __inline __m512d |
1648 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1649 | _mm512_cvt_roundepu64_pd (__m512i __A, const int __R) |
1650 | { |
1651 | return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, |
1652 | (__v8df) |
1653 | _mm512_setzero_pd (), |
1654 | (__mmask8) -1, |
1655 | __R); |
1656 | } |
1657 | |
1658 | extern __inline __m512d |
1659 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1660 | _mm512_mask_cvt_roundepu64_pd (__m512d __W, __mmask8 __U, __m512i __A, |
1661 | const int __R) |
1662 | { |
1663 | return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, |
1664 | (__v8df) __W, |
1665 | (__mmask8) __U, |
1666 | __R); |
1667 | } |
1668 | |
1669 | extern __inline __m512d |
1670 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1671 | _mm512_maskz_cvt_roundepu64_pd (__mmask8 __U, __m512i __A, |
1672 | const int __R) |
1673 | { |
1674 | return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, |
1675 | (__v8df) |
1676 | _mm512_setzero_pd (), |
1677 | (__mmask8) __U, |
1678 | __R); |
1679 | } |
1680 | |
1681 | extern __inline __m512d |
1682 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1683 | _mm512_reduce_pd (__m512d __A, int __B) |
1684 | { |
1685 | return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, |
1686 | (__v8df) |
1687 | _mm512_setzero_pd (), |
1688 | (__mmask8) -1); |
1689 | } |
1690 | |
1691 | extern __inline __m512d |
1692 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1693 | _mm512_mask_reduce_pd (__m512d __W, __mmask8 __U, __m512d __A, int __B) |
1694 | { |
1695 | return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, |
1696 | (__v8df) __W, |
1697 | (__mmask8) __U); |
1698 | } |
1699 | |
1700 | extern __inline __m512d |
1701 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1702 | _mm512_maskz_reduce_pd (__mmask8 __U, __m512d __A, int __B) |
1703 | { |
1704 | return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, |
1705 | (__v8df) |
1706 | _mm512_setzero_pd (), |
1707 | (__mmask8) __U); |
1708 | } |
1709 | |
1710 | extern __inline __m512 |
1711 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1712 | _mm512_reduce_ps (__m512 __A, int __B) |
1713 | { |
1714 | return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, |
1715 | (__v16sf) |
1716 | _mm512_setzero_ps (), |
1717 | (__mmask16) -1); |
1718 | } |
1719 | |
1720 | extern __inline __m512 |
1721 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1722 | _mm512_mask_reduce_ps (__m512 __W, __mmask16 __U, __m512 __A, int __B) |
1723 | { |
1724 | return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, |
1725 | (__v16sf) __W, |
1726 | (__mmask16) __U); |
1727 | } |
1728 | |
1729 | extern __inline __m512 |
1730 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1731 | _mm512_maskz_reduce_ps (__mmask16 __U, __m512 __A, int __B) |
1732 | { |
1733 | return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, |
1734 | (__v16sf) |
1735 | _mm512_setzero_ps (), |
1736 | (__mmask16) __U); |
1737 | } |
1738 | |
1739 | extern __inline __m256 |
1740 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1741 | _mm512_extractf32x8_ps (__m512 __A, const int __imm) |
1742 | { |
1743 | return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, |
1744 | __imm, |
1745 | (__v8sf) |
1746 | _mm256_setzero_ps (), |
1747 | (__mmask8) -1); |
1748 | } |
1749 | |
1750 | extern __inline __m256 |
1751 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1752 | _mm512_mask_extractf32x8_ps (__m256 __W, __mmask8 __U, __m512 __A, |
1753 | const int __imm) |
1754 | { |
1755 | return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, |
1756 | __imm, |
1757 | (__v8sf) __W, |
1758 | (__mmask8) __U); |
1759 | } |
1760 | |
1761 | extern __inline __m256 |
1762 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1763 | _mm512_maskz_extractf32x8_ps (__mmask8 __U, __m512 __A, |
1764 | const int __imm) |
1765 | { |
1766 | return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, |
1767 | __imm, |
1768 | (__v8sf) |
1769 | _mm256_setzero_ps (), |
1770 | (__mmask8) __U); |
1771 | } |
1772 | |
1773 | extern __inline __m128d |
1774 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1775 | _mm512_extractf64x2_pd (__m512d __A, const int __imm) |
1776 | { |
1777 | return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, |
1778 | __imm, |
1779 | (__v2df) |
1780 | _mm_setzero_pd (), |
1781 | (__mmask8) -1); |
1782 | } |
1783 | |
1784 | extern __inline __m128d |
1785 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1786 | _mm512_mask_extractf64x2_pd (__m128d __W, __mmask8 __U, __m512d __A, |
1787 | const int __imm) |
1788 | { |
1789 | return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, |
1790 | __imm, |
1791 | (__v2df) __W, |
1792 | (__mmask8) |
1793 | __U); |
1794 | } |
1795 | |
1796 | extern __inline __m128d |
1797 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1798 | _mm512_maskz_extractf64x2_pd (__mmask8 __U, __m512d __A, |
1799 | const int __imm) |
1800 | { |
1801 | return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, |
1802 | __imm, |
1803 | (__v2df) |
1804 | _mm_setzero_pd (), |
1805 | (__mmask8) |
1806 | __U); |
1807 | } |
1808 | |
1809 | extern __inline __m256i |
1810 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1811 | _mm512_extracti32x8_epi32 (__m512i __A, const int __imm) |
1812 | { |
1813 | return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, |
1814 | __imm, |
1815 | (__v8si) |
1816 | _mm256_setzero_si256 (), |
1817 | (__mmask8) -1); |
1818 | } |
1819 | |
1820 | extern __inline __m256i |
1821 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1822 | _mm512_mask_extracti32x8_epi32 (__m256i __W, __mmask8 __U, __m512i __A, |
1823 | const int __imm) |
1824 | { |
1825 | return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, |
1826 | __imm, |
1827 | (__v8si) __W, |
1828 | (__mmask8) __U); |
1829 | } |
1830 | |
1831 | extern __inline __m256i |
1832 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1833 | _mm512_maskz_extracti32x8_epi32 (__mmask8 __U, __m512i __A, |
1834 | const int __imm) |
1835 | { |
1836 | return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, |
1837 | __imm, |
1838 | (__v8si) |
1839 | _mm256_setzero_si256 (), |
1840 | (__mmask8) __U); |
1841 | } |
1842 | |
1843 | extern __inline __m128i |
1844 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1845 | _mm512_extracti64x2_epi64 (__m512i __A, const int __imm) |
1846 | { |
1847 | return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, |
1848 | __imm, |
1849 | (__v2di) |
1850 | _mm_setzero_si128 (), |
1851 | (__mmask8) -1); |
1852 | } |
1853 | |
1854 | extern __inline __m128i |
1855 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1856 | _mm512_mask_extracti64x2_epi64 (__m128i __W, __mmask8 __U, __m512i __A, |
1857 | const int __imm) |
1858 | { |
1859 | return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, |
1860 | __imm, |
1861 | (__v2di) __W, |
1862 | (__mmask8) |
1863 | __U); |
1864 | } |
1865 | |
1866 | extern __inline __m128i |
1867 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1868 | _mm512_maskz_extracti64x2_epi64 (__mmask8 __U, __m512i __A, |
1869 | const int __imm) |
1870 | { |
1871 | return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, |
1872 | __imm, |
1873 | (__v2di) |
1874 | _mm_setzero_si128 (), |
1875 | (__mmask8) |
1876 | __U); |
1877 | } |
1878 | |
1879 | extern __inline __m512d |
1880 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1881 | _mm512_range_round_pd (__m512d __A, __m512d __B, int __C, |
1882 | const int __R) |
1883 | { |
1884 | return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, |
1885 | (__v8df) __B, __C, |
1886 | (__v8df) |
1887 | _mm512_setzero_pd (), |
1888 | (__mmask8) -1, |
1889 | __R); |
1890 | } |
1891 | |
1892 | extern __inline __m512d |
1893 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1894 | _mm512_mask_range_round_pd (__m512d __W, __mmask8 __U, |
1895 | __m512d __A, __m512d __B, int __C, |
1896 | const int __R) |
1897 | { |
1898 | return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, |
1899 | (__v8df) __B, __C, |
1900 | (__v8df) __W, |
1901 | (__mmask8) __U, |
1902 | __R); |
1903 | } |
1904 | |
1905 | extern __inline __m512d |
1906 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1907 | _mm512_maskz_range_round_pd (__mmask8 __U, __m512d __A, __m512d __B, |
1908 | int __C, const int __R) |
1909 | { |
1910 | return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, |
1911 | (__v8df) __B, __C, |
1912 | (__v8df) |
1913 | _mm512_setzero_pd (), |
1914 | (__mmask8) __U, |
1915 | __R); |
1916 | } |
1917 | |
1918 | extern __inline __m512 |
1919 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1920 | _mm512_range_round_ps (__m512 __A, __m512 __B, int __C, const int __R) |
1921 | { |
1922 | return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, |
1923 | (__v16sf) __B, __C, |
1924 | (__v16sf) |
1925 | _mm512_setzero_ps (), |
1926 | (__mmask16) -1, |
1927 | __R); |
1928 | } |
1929 | |
1930 | extern __inline __m512 |
1931 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1932 | _mm512_mask_range_round_ps (__m512 __W, __mmask16 __U, |
1933 | __m512 __A, __m512 __B, int __C, |
1934 | const int __R) |
1935 | { |
1936 | return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, |
1937 | (__v16sf) __B, __C, |
1938 | (__v16sf) __W, |
1939 | (__mmask16) __U, |
1940 | __R); |
1941 | } |
1942 | |
1943 | extern __inline __m512 |
1944 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1945 | _mm512_maskz_range_round_ps (__mmask16 __U, __m512 __A, __m512 __B, |
1946 | int __C, const int __R) |
1947 | { |
1948 | return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, |
1949 | (__v16sf) __B, __C, |
1950 | (__v16sf) |
1951 | _mm512_setzero_ps (), |
1952 | (__mmask16) __U, |
1953 | __R); |
1954 | } |
1955 | |
1956 | extern __inline __m512i |
1957 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1958 | _mm512_inserti32x8 (__m512i __A, __m256i __B, const int __imm) |
1959 | { |
1960 | return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, |
1961 | (__v8si) __B, |
1962 | __imm, |
1963 | (__v16si) |
1964 | _mm512_setzero_si512 (), |
1965 | (__mmask16) -1); |
1966 | } |
1967 | |
1968 | extern __inline __m512i |
1969 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1970 | _mm512_mask_inserti32x8 (__m512i __W, __mmask16 __U, __m512i __A, |
1971 | __m256i __B, const int __imm) |
1972 | { |
1973 | return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, |
1974 | (__v8si) __B, |
1975 | __imm, |
1976 | (__v16si) __W, |
1977 | (__mmask16) __U); |
1978 | } |
1979 | |
1980 | extern __inline __m512i |
1981 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1982 | _mm512_maskz_inserti32x8 (__mmask16 __U, __m512i __A, __m256i __B, |
1983 | const int __imm) |
1984 | { |
1985 | return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, |
1986 | (__v8si) __B, |
1987 | __imm, |
1988 | (__v16si) |
1989 | _mm512_setzero_si512 (), |
1990 | (__mmask16) __U); |
1991 | } |
1992 | |
1993 | extern __inline __m512 |
1994 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
1995 | _mm512_insertf32x8 (__m512 __A, __m256 __B, const int __imm) |
1996 | { |
1997 | return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, |
1998 | (__v8sf) __B, |
1999 | __imm, |
2000 | (__v16sf) |
2001 | _mm512_setzero_ps (), |
2002 | (__mmask16) -1); |
2003 | } |
2004 | |
2005 | extern __inline __m512 |
2006 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2007 | _mm512_mask_insertf32x8 (__m512 __W, __mmask16 __U, __m512 __A, |
2008 | __m256 __B, const int __imm) |
2009 | { |
2010 | return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, |
2011 | (__v8sf) __B, |
2012 | __imm, |
2013 | (__v16sf) __W, |
2014 | (__mmask16) __U); |
2015 | } |
2016 | |
2017 | extern __inline __m512 |
2018 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2019 | _mm512_maskz_insertf32x8 (__mmask16 __U, __m512 __A, __m256 __B, |
2020 | const int __imm) |
2021 | { |
2022 | return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, |
2023 | (__v8sf) __B, |
2024 | __imm, |
2025 | (__v16sf) |
2026 | _mm512_setzero_ps (), |
2027 | (__mmask16) __U); |
2028 | } |
2029 | |
2030 | extern __inline __m512i |
2031 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2032 | _mm512_inserti64x2 (__m512i __A, __m128i __B, const int __imm) |
2033 | { |
2034 | return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, |
2035 | (__v2di) __B, |
2036 | __imm, |
2037 | (__v8di) |
2038 | _mm512_setzero_si512 (), |
2039 | (__mmask8) -1); |
2040 | } |
2041 | |
2042 | extern __inline __m512i |
2043 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2044 | _mm512_mask_inserti64x2 (__m512i __W, __mmask8 __U, __m512i __A, |
2045 | __m128i __B, const int __imm) |
2046 | { |
2047 | return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, |
2048 | (__v2di) __B, |
2049 | __imm, |
2050 | (__v8di) __W, |
2051 | (__mmask8) |
2052 | __U); |
2053 | } |
2054 | |
2055 | extern __inline __m512i |
2056 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2057 | _mm512_maskz_inserti64x2 (__mmask8 __U, __m512i __A, __m128i __B, |
2058 | const int __imm) |
2059 | { |
2060 | return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, |
2061 | (__v2di) __B, |
2062 | __imm, |
2063 | (__v8di) |
2064 | _mm512_setzero_si512 (), |
2065 | (__mmask8) |
2066 | __U); |
2067 | } |
2068 | |
2069 | extern __inline __m512d |
2070 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2071 | _mm512_insertf64x2 (__m512d __A, __m128d __B, const int __imm) |
2072 | { |
2073 | return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, |
2074 | (__v2df) __B, |
2075 | __imm, |
2076 | (__v8df) |
2077 | _mm512_setzero_pd (), |
2078 | (__mmask8) -1); |
2079 | } |
2080 | |
2081 | extern __inline __m512d |
2082 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2083 | _mm512_mask_insertf64x2 (__m512d __W, __mmask8 __U, __m512d __A, |
2084 | __m128d __B, const int __imm) |
2085 | { |
2086 | return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, |
2087 | (__v2df) __B, |
2088 | __imm, |
2089 | (__v8df) __W, |
2090 | (__mmask8) |
2091 | __U); |
2092 | } |
2093 | |
2094 | extern __inline __m512d |
2095 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2096 | _mm512_maskz_insertf64x2 (__mmask8 __U, __m512d __A, __m128d __B, |
2097 | const int __imm) |
2098 | { |
2099 | return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, |
2100 | (__v2df) __B, |
2101 | __imm, |
2102 | (__v8df) |
2103 | _mm512_setzero_pd (), |
2104 | (__mmask8) |
2105 | __U); |
2106 | } |
2107 | |
2108 | extern __inline __mmask8 |
2109 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2110 | _mm512_mask_fpclass_pd_mask (__mmask8 __U, __m512d __A, |
2111 | const int __imm) |
2112 | { |
2113 | return (__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) __A, |
2114 | __imm, __U); |
2115 | } |
2116 | |
2117 | extern __inline __mmask8 |
2118 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2119 | _mm512_fpclass_pd_mask (__m512d __A, const int __imm) |
2120 | { |
2121 | return (__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) __A, |
2122 | __imm, |
2123 | (__mmask8) -1); |
2124 | } |
2125 | |
2126 | extern __inline __mmask16 |
2127 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2128 | _mm512_mask_fpclass_ps_mask (__mmask16 __U, __m512 __A, |
2129 | const int __imm) |
2130 | { |
2131 | return (__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) __A, |
2132 | __imm, __U); |
2133 | } |
2134 | |
2135 | extern __inline __mmask16 |
2136 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
2137 | _mm512_fpclass_ps_mask (__m512 __A, const int __imm) |
2138 | { |
2139 | return (__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) __A, |
2140 | __imm, |
2141 | (__mmask16) -1); |
2142 | } |
2143 | |
2144 | #else |
2145 | #define _kshiftli_mask8(X, Y) \ |
2146 | ((__mmask8) __builtin_ia32_kshiftliqi ((__mmask8)(X), (__mmask8)(Y))) |
2147 | |
2148 | #define _kshiftri_mask8(X, Y) \ |
2149 | ((__mmask8) __builtin_ia32_kshiftriqi ((__mmask8)(X), (__mmask8)(Y))) |
2150 | |
2151 | #define _mm_range_sd(A, B, C) \ |
2152 | ((__m128d) __builtin_ia32_rangesd128_round ((__v2df)(__m128d)(A), \ |
2153 | (__v2df)(__m128d)(B), (int)(C), \ |
2154 | _MM_FROUND_CUR_DIRECTION)) |
2155 | |
2156 | #define _mm_range_ss(A, B, C) \ |
2157 | ((__m128) __builtin_ia32_rangess128_round ((__v4sf)(__m128)(A), \ |
2158 | (__v4sf)(__m128)(B), (int)(C), \ |
2159 | _MM_FROUND_CUR_DIRECTION)) |
2160 | |
2161 | #define _mm_range_round_sd(A, B, C, R) \ |
2162 | ((__m128d) __builtin_ia32_rangesd128_round ((__v2df)(__m128d)(A), \ |
2163 | (__v2df)(__m128d)(B), (int)(C), (R))) |
2164 | |
2165 | #define _mm_range_round_ss(A, B, C, R) \ |
2166 | ((__m128) __builtin_ia32_rangess128_round ((__v4sf)(__m128)(A), \ |
2167 | (__v4sf)(__m128)(B), (int)(C), (R))) |
2168 | |
2169 | #define _mm512_cvtt_roundpd_epi64(A, B) \ |
2170 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di) \ |
2171 | _mm512_setzero_si512 (), \ |
2172 | -1, (B))) |
2173 | |
2174 | #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, B) \ |
2175 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di)(W), (U), (B))) |
2176 | |
2177 | #define _mm512_maskz_cvtt_roundpd_epi64(U, A, B) \ |
2178 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2179 | |
2180 | #define _mm512_cvtt_roundpd_epu64(A, B) \ |
2181 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2182 | |
2183 | #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, B) \ |
2184 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)(W), (U), (B))) |
2185 | |
2186 | #define _mm512_maskz_cvtt_roundpd_epu64(U, A, B) \ |
2187 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2188 | |
2189 | #define _mm512_cvtt_roundps_epi64(A, B) \ |
2190 | ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2191 | |
2192 | #define _mm512_mask_cvtt_roundps_epi64(W, U, A, B) \ |
2193 | ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)(W), (U), (B))) |
2194 | |
2195 | #define _mm512_maskz_cvtt_roundps_epi64(U, A, B) \ |
2196 | ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2197 | |
2198 | #define _mm512_cvtt_roundps_epu64(A, B) \ |
2199 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2200 | |
2201 | #define _mm512_mask_cvtt_roundps_epu64(W, U, A, B) \ |
2202 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)(W), (U), (B))) |
2203 | |
2204 | #define _mm512_maskz_cvtt_roundps_epu64(U, A, B) \ |
2205 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2206 | |
2207 | #define _mm512_cvt_roundpd_epi64(A, B) \ |
2208 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2209 | |
2210 | #define _mm512_mask_cvt_roundpd_epi64(W, U, A, B) \ |
2211 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)(W), (U), (B))) |
2212 | |
2213 | #define _mm512_maskz_cvt_roundpd_epi64(U, A, B) \ |
2214 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2215 | |
2216 | #define _mm512_cvt_roundpd_epu64(A, B) \ |
2217 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2218 | |
2219 | #define _mm512_mask_cvt_roundpd_epu64(W, U, A, B) \ |
2220 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)(W), (U), (B))) |
2221 | |
2222 | #define _mm512_maskz_cvt_roundpd_epu64(U, A, B) \ |
2223 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2224 | |
2225 | #define _mm512_cvt_roundps_epi64(A, B) \ |
2226 | ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2227 | |
2228 | #define _mm512_mask_cvt_roundps_epi64(W, U, A, B) \ |
2229 | ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)(W), (U), (B))) |
2230 | |
2231 | #define _mm512_maskz_cvt_roundps_epi64(U, A, B) \ |
2232 | ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2233 | |
2234 | #define _mm512_cvt_roundps_epu64(A, B) \ |
2235 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) |
2236 | |
2237 | #define _mm512_mask_cvt_roundps_epu64(W, U, A, B) \ |
2238 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)(W), (U), (B))) |
2239 | |
2240 | #define _mm512_maskz_cvt_roundps_epu64(U, A, B) \ |
2241 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) |
2242 | |
2243 | #define _mm512_cvt_roundepi64_ps(A, B) \ |
2244 | ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), -1, (B))) |
2245 | |
2246 | #define _mm512_mask_cvt_roundepi64_ps(W, U, A, B) \ |
2247 | ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (W), (U), (B))) |
2248 | |
2249 | #define _mm512_maskz_cvt_roundepi64_ps(U, A, B) \ |
2250 | ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), (U), (B))) |
2251 | |
2252 | #define _mm512_cvt_roundepu64_ps(A, B) \ |
2253 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), -1, (B))) |
2254 | |
2255 | #define _mm512_mask_cvt_roundepu64_ps(W, U, A, B) \ |
2256 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (W), (U), (B))) |
2257 | |
2258 | #define _mm512_maskz_cvt_roundepu64_ps(U, A, B) \ |
2259 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), (U), (B))) |
2260 | |
2261 | #define _mm512_cvt_roundepi64_pd(A, B) \ |
2262 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), -1, (B))) |
2263 | |
2264 | #define _mm512_mask_cvt_roundepi64_pd(W, U, A, B) \ |
2265 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (W), (U), (B))) |
2266 | |
2267 | #define _mm512_maskz_cvt_roundepi64_pd(U, A, B) \ |
2268 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), (U), (B))) |
2269 | |
2270 | #define _mm512_cvt_roundepu64_pd(A, B) \ |
2271 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), -1, (B))) |
2272 | |
2273 | #define _mm512_mask_cvt_roundepu64_pd(W, U, A, B) \ |
2274 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (W), (U), (B))) |
2275 | |
2276 | #define _mm512_maskz_cvt_roundepu64_pd(U, A, B) \ |
2277 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), (U), (B))) |
2278 | |
2279 | #define _mm512_reduce_pd(A, B) \ |
2280 | ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ |
2281 | (int)(B), (__v8df)_mm512_setzero_pd (), (__mmask8)-1)) |
2282 | |
2283 | #define _mm512_mask_reduce_pd(W, U, A, B) \ |
2284 | ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ |
2285 | (int)(B), (__v8df)(__m512d)(W), (__mmask8)(U))) |
2286 | |
2287 | #define _mm512_maskz_reduce_pd(U, A, B) \ |
2288 | ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ |
2289 | (int)(B), (__v8df)_mm512_setzero_pd (), (__mmask8)(U))) |
2290 | |
2291 | #define _mm512_reduce_ps(A, B) \ |
2292 | ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ |
2293 | (int)(B), (__v16sf)_mm512_setzero_ps (), (__mmask16)-1)) |
2294 | |
2295 | #define _mm512_mask_reduce_ps(W, U, A, B) \ |
2296 | ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ |
2297 | (int)(B), (__v16sf)(__m512)(W), (__mmask16)(U))) |
2298 | |
2299 | #define _mm512_maskz_reduce_ps(U, A, B) \ |
2300 | ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ |
2301 | (int)(B), (__v16sf)_mm512_setzero_ps (), (__mmask16)(U))) |
2302 | |
2303 | #define _mm512_extractf32x8_ps(X, C) \ |
2304 | ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ |
2305 | (int) (C), (__v8sf)(__m256) _mm256_setzero_ps (), (__mmask8)-1)) |
2306 | |
2307 | #define _mm512_mask_extractf32x8_ps(W, U, X, C) \ |
2308 | ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ |
2309 | (int) (C), (__v8sf)(__m256) (W), (__mmask8) (U))) |
2310 | |
2311 | #define _mm512_maskz_extractf32x8_ps(U, X, C) \ |
2312 | ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ |
2313 | (int) (C), (__v8sf)(__m256) _mm256_setzero_ps (), (__mmask8) (U))) |
2314 | |
2315 | #define _mm512_extractf64x2_pd(X, C) \ |
2316 | ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ |
2317 | (int) (C), (__v2df)(__m128d) _mm_setzero_pd (), (__mmask8)-1)) |
2318 | |
2319 | #define _mm512_mask_extractf64x2_pd(W, U, X, C) \ |
2320 | ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ |
2321 | (int) (C), (__v2df)(__m128d) (W), (__mmask8) (U))) |
2322 | |
2323 | #define _mm512_maskz_extractf64x2_pd(U, X, C) \ |
2324 | ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ |
2325 | (int) (C), (__v2df)(__m128d) _mm_setzero_pd (), (__mmask8) (U))) |
2326 | |
2327 | #define _mm512_extracti32x8_epi32(X, C) \ |
2328 | ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ |
2329 | (int) (C), (__v8si)(__m256i) _mm256_setzero_si256 (), (__mmask8)-1)) |
2330 | |
2331 | #define _mm512_mask_extracti32x8_epi32(W, U, X, C) \ |
2332 | ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ |
2333 | (int) (C), (__v8si)(__m256i) (W), (__mmask8) (U))) |
2334 | |
2335 | #define _mm512_maskz_extracti32x8_epi32(U, X, C) \ |
2336 | ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ |
2337 | (int) (C), (__v8si)(__m256i) _mm256_setzero_si256 (), (__mmask8) (U))) |
2338 | |
2339 | #define _mm512_extracti64x2_epi64(X, C) \ |
2340 | ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ |
2341 | (int) (C), (__v2di)(__m128i) _mm_setzero_si128 (), (__mmask8)-1)) |
2342 | |
2343 | #define _mm512_mask_extracti64x2_epi64(W, U, X, C) \ |
2344 | ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ |
2345 | (int) (C), (__v2di)(__m128i) (W), (__mmask8) (U))) |
2346 | |
2347 | #define _mm512_maskz_extracti64x2_epi64(U, X, C) \ |
2348 | ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ |
2349 | (int) (C), (__v2di)(__m128i) _mm_setzero_si128 (), (__mmask8) (U))) |
2350 | |
2351 | #define _mm512_range_pd(A, B, C) \ |
2352 | ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ |
2353 | (__v8df)(__m512d)(B), (int)(C), \ |
2354 | (__v8df)_mm512_setzero_pd (), (__mmask8)-1, _MM_FROUND_CUR_DIRECTION)) |
2355 | |
2356 | #define _mm512_mask_range_pd(W, U, A, B, C) \ |
2357 | ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ |
2358 | (__v8df)(__m512d)(B), (int)(C), \ |
2359 | (__v8df)(__m512d)(W), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) |
2360 | |
2361 | #define _mm512_maskz_range_pd(U, A, B, C) \ |
2362 | ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ |
2363 | (__v8df)(__m512d)(B), (int)(C), \ |
2364 | (__v8df)_mm512_setzero_pd (), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) |
2365 | |
2366 | #define _mm512_range_ps(A, B, C) \ |
2367 | ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ |
2368 | (__v16sf)(__m512)(B), (int)(C), \ |
2369 | (__v16sf)_mm512_setzero_ps (), (__mmask16)-1, _MM_FROUND_CUR_DIRECTION)) |
2370 | |
2371 | #define _mm512_mask_range_ps(W, U, A, B, C) \ |
2372 | ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ |
2373 | (__v16sf)(__m512)(B), (int)(C), \ |
2374 | (__v16sf)(__m512)(W), (__mmask16)(U), _MM_FROUND_CUR_DIRECTION)) |
2375 | |
2376 | #define _mm512_maskz_range_ps(U, A, B, C) \ |
2377 | ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ |
2378 | (__v16sf)(__m512)(B), (int)(C), \ |
2379 | (__v16sf)_mm512_setzero_ps (), (__mmask16)(U), _MM_FROUND_CUR_DIRECTION)) |
2380 | |
2381 | #define _mm512_range_round_pd(A, B, C, R) \ |
2382 | ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ |
2383 | (__v8df)(__m512d)(B), (int)(C), \ |
2384 | (__v8df)_mm512_setzero_pd (), (__mmask8)-1, (R))) |
2385 | |
2386 | #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ |
2387 | ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ |
2388 | (__v8df)(__m512d)(B), (int)(C), \ |
2389 | (__v8df)(__m512d)(W), (__mmask8)(U), (R))) |
2390 | |
2391 | #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ |
2392 | ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ |
2393 | (__v8df)(__m512d)(B), (int)(C), \ |
2394 | (__v8df)_mm512_setzero_pd (), (__mmask8)(U), (R))) |
2395 | |
2396 | #define _mm512_range_round_ps(A, B, C, R) \ |
2397 | ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ |
2398 | (__v16sf)(__m512)(B), (int)(C), \ |
2399 | (__v16sf)_mm512_setzero_ps (), (__mmask16)-1, (R))) |
2400 | |
2401 | #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ |
2402 | ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ |
2403 | (__v16sf)(__m512)(B), (int)(C), \ |
2404 | (__v16sf)(__m512)(W), (__mmask16)(U), (R))) |
2405 | |
2406 | #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ |
2407 | ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ |
2408 | (__v16sf)(__m512)(B), (int)(C), \ |
2409 | (__v16sf)_mm512_setzero_ps (), (__mmask16)(U), (R))) |
2410 | |
2411 | #define _mm512_insertf64x2(X, Y, C) \ |
2412 | ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ |
2413 | (__v2df)(__m128d) (Y), (int) (C), (__v8df)(__m512d) (X), \ |
2414 | (__mmask8)-1)) |
2415 | |
2416 | #define _mm512_mask_insertf64x2(W, U, X, Y, C) \ |
2417 | ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ |
2418 | (__v2df)(__m128d) (Y), (int) (C), (__v8df)(__m512d) (W), \ |
2419 | (__mmask8) (U))) |
2420 | |
2421 | #define _mm512_maskz_insertf64x2(U, X, Y, C) \ |
2422 | ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ |
2423 | (__v2df)(__m128d) (Y), (int) (C), \ |
2424 | (__v8df)(__m512d) _mm512_setzero_pd (), (__mmask8) (U))) |
2425 | |
2426 | #define _mm512_inserti64x2(X, Y, C) \ |
2427 | ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ |
2428 | (__v2di)(__m128i) (Y), (int) (C), (__v8di)(__m512i) (X), (__mmask8)-1)) |
2429 | |
2430 | #define _mm512_mask_inserti64x2(W, U, X, Y, C) \ |
2431 | ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ |
2432 | (__v2di)(__m128i) (Y), (int) (C), (__v8di)(__m512i) (W), \ |
2433 | (__mmask8) (U))) |
2434 | |
2435 | #define _mm512_maskz_inserti64x2(U, X, Y, C) \ |
2436 | ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ |
2437 | (__v2di)(__m128i) (Y), (int) (C), \ |
2438 | (__v8di)(__m512i) _mm512_setzero_si512 (), (__mmask8) (U))) |
2439 | |
2440 | #define _mm512_insertf32x8(X, Y, C) \ |
2441 | ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ |
2442 | (__v8sf)(__m256) (Y), (int) (C),\ |
2443 | (__v16sf)(__m512)_mm512_setzero_ps (),\ |
2444 | (__mmask16)-1)) |
2445 | |
2446 | #define _mm512_mask_insertf32x8(W, U, X, Y, C) \ |
2447 | ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ |
2448 | (__v8sf)(__m256) (Y), (int) (C),\ |
2449 | (__v16sf)(__m512)(W),\ |
2450 | (__mmask16)(U))) |
2451 | |
2452 | #define _mm512_maskz_insertf32x8(U, X, Y, C) \ |
2453 | ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ |
2454 | (__v8sf)(__m256) (Y), (int) (C),\ |
2455 | (__v16sf)(__m512)_mm512_setzero_ps (),\ |
2456 | (__mmask16)(U))) |
2457 | |
2458 | #define _mm512_inserti32x8(X, Y, C) \ |
2459 | ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ |
2460 | (__v8si)(__m256i) (Y), (int) (C),\ |
2461 | (__v16si)(__m512i)_mm512_setzero_si512 (),\ |
2462 | (__mmask16)-1)) |
2463 | |
2464 | #define _mm512_mask_inserti32x8(W, U, X, Y, C) \ |
2465 | ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ |
2466 | (__v8si)(__m256i) (Y), (int) (C),\ |
2467 | (__v16si)(__m512i)(W),\ |
2468 | (__mmask16)(U))) |
2469 | |
2470 | #define _mm512_maskz_inserti32x8(U, X, Y, C) \ |
2471 | ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ |
2472 | (__v8si)(__m256i) (Y), (int) (C),\ |
2473 | (__v16si)(__m512i)_mm512_setzero_si512 (),\ |
2474 | (__mmask16)(U))) |
2475 | |
2476 | #define _mm_fpclass_ss_mask(X, C) \ |
2477 | ((__mmask8) __builtin_ia32_fpclassss ((__v4sf) (__m128) (X), (int) (C))) \ |
2478 | |
2479 | #define _mm_fpclass_sd_mask(X, C) \ |
2480 | ((__mmask8) __builtin_ia32_fpclasssd ((__v2df) (__m128d) (X), (int) (C))) \ |
2481 | |
2482 | #define _mm512_mask_fpclass_pd_mask(u, X, C) \ |
2483 | ((__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) (__m512d) (X), \ |
2484 | (int) (C), (__mmask8)(u))) |
2485 | |
2486 | #define _mm512_mask_fpclass_ps_mask(u, x, c) \ |
2487 | ((__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) (__m512) (x),\ |
2488 | (int) (c),(__mmask8)(u))) |
2489 | |
2490 | #define _mm512_fpclass_pd_mask(X, C) \ |
2491 | ((__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) (__m512d) (X), \ |
2492 | (int) (C), (__mmask8)-1)) |
2493 | |
2494 | #define _mm512_fpclass_ps_mask(x, c) \ |
2495 | ((__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) (__m512) (x),\ |
2496 | (int) (c),(__mmask8)-1)) |
2497 | |
2498 | #define _mm_reduce_sd(A, B, C) \ |
2499 | ((__m128d) __builtin_ia32_reducesd_mask ((__v2df)(__m128d)(A), \ |
2500 | (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ |
2501 | (__mmask8)-1)) |
2502 | |
2503 | #define _mm_mask_reduce_sd(W, U, A, B, C) \ |
2504 | ((__m128d) __builtin_ia32_reducesd_mask ((__v2df)(__m128d)(A), \ |
2505 | (__v2df)(__m128d)(B), (int)(C), (__v2df)(__m128d)(W), (__mmask8)(U))) |
2506 | |
2507 | #define _mm_maskz_reduce_sd(U, A, B, C) \ |
2508 | ((__m128d) __builtin_ia32_reducesd_mask ((__v2df)(__m128d)(A), \ |
2509 | (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ |
2510 | (__mmask8)(U))) |
2511 | |
2512 | #define _mm_reduce_ss(A, B, C) \ |
2513 | ((__m128) __builtin_ia32_reducess_mask ((__v4sf)(__m128)(A), \ |
2514 | (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ |
2515 | (__mmask8)-1)) |
2516 | |
2517 | #define _mm_mask_reduce_ss(W, U, A, B, C) \ |
2518 | ((__m128) __builtin_ia32_reducess_mask ((__v4sf)(__m128)(A), \ |
2519 | (__v4sf)(__m128)(B), (int)(C), (__v4sf)(__m128)(W), (__mmask8)(U))) |
2520 | |
2521 | #define _mm_maskz_reduce_ss(U, A, B, C) \ |
2522 | ((__m128) __builtin_ia32_reducess_mask ((__v4sf)(__m128)(A), \ |
2523 | (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ |
2524 | (__mmask8)(U))) |
2525 | |
2526 | |
2527 | |
2528 | #endif |
2529 | |
2530 | #ifdef __DISABLE_AVX512DQ__ |
2531 | #undef __DISABLE_AVX512DQ__ |
2532 | #pragma GCC pop_options |
2533 | #endif /* __DISABLE_AVX512DQ__ */ |
2534 | |
2535 | #endif /* _AVX512DQINTRIN_H_INCLUDED */ |
2536 |
Warning: That file was not part of the compilation database. It may have many parsing errors.