]>
Commit | Line | Data |
---|---|---|
cbe34bb5 | 1 | /* Copyright (C) 2014-2017 Free Software Foundation, Inc. |
936c0fe4 AI |
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 <avx512bwintrin.h> directly; include <immintrin.h> instead." | |
26 | #endif | |
27 | ||
28 | #ifndef _AVX512BWINTRIN_H_INCLUDED | |
29 | #define _AVX512BWINTRIN_H_INCLUDED | |
30 | ||
31 | #ifndef __AVX512BW__ | |
32 | #pragma GCC push_options | |
33 | #pragma GCC target("avx512bw") | |
34 | #define __DISABLE_AVX512BW__ | |
35 | #endif /* __AVX512BW__ */ | |
36 | ||
37 | /* Internal data types for implementing the intrinsics. */ | |
38 | typedef short __v32hi __attribute__ ((__vector_size__ (64))); | |
39 | typedef char __v64qi __attribute__ ((__vector_size__ (64))); | |
40 | ||
41 | typedef unsigned long long __mmask64; | |
42 | ||
75d727d2 AS |
43 | extern __inline __mmask32 |
44 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
45 | _kadd_mask32 (__mmask32 __A, __mmask32 __B) | |
46 | { | |
47 | return (__mmask32) __builtin_ia32_kaddsi ((__mmask32) __A, (__mmask32) __B); | |
48 | } | |
49 | ||
50 | extern __inline __mmask64 | |
51 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
52 | _kadd_mask64 (__mmask64 __A, __mmask64 __B) | |
53 | { | |
54 | return (__mmask64) __builtin_ia32_kadddi ((__mmask64) __A, (__mmask64) __B); | |
55 | } | |
56 | ||
7cdb6e4c AS |
57 | extern __inline unsigned int |
58 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
59 | _cvtmask32_u32 (__mmask32 __A) | |
60 | { | |
61 | return (unsigned int) __builtin_ia32_kmovd ((__mmask32) __A); | |
62 | } | |
63 | ||
64 | extern __inline unsigned long long | |
65 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
66 | _cvtmask64_u64 (__mmask64 __A) | |
67 | { | |
68 | return (unsigned long long) __builtin_ia32_kmovq ((__mmask64) __A); | |
69 | } | |
70 | ||
71 | extern __inline __mmask32 | |
72 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
73 | _cvtu32_mask32 (unsigned int __A) | |
74 | { | |
75 | return (__mmask32) __builtin_ia32_kmovd ((__mmask32) __A); | |
76 | } | |
77 | ||
78 | extern __inline __mmask64 | |
79 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
80 | _cvtu64_mask64 (unsigned long long __A) | |
81 | { | |
82 | return (__mmask64) __builtin_ia32_kmovq ((__mmask64) __A); | |
83 | } | |
84 | ||
85 | extern __inline __mmask32 | |
86 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
87 | _load_mask32 (__mmask32 *__A) | |
88 | { | |
89 | return (__mmask32) __builtin_ia32_kmovd (*__A); | |
90 | } | |
91 | ||
92 | extern __inline __mmask64 | |
93 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
94 | _load_mask64 (__mmask64 *__A) | |
95 | { | |
96 | return (__mmask64) __builtin_ia32_kmovq (*(__mmask64 *) __A); | |
97 | } | |
98 | ||
99 | extern __inline void | |
100 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
101 | _store_mask32 (__mmask32 *__A, __mmask32 __B) | |
102 | { | |
103 | *(__mmask32 *) __A = __builtin_ia32_kmovd (__B); | |
104 | } | |
105 | ||
106 | extern __inline void | |
107 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
108 | _store_mask64 (__mmask64 *__A, __mmask64 __B) | |
109 | { | |
110 | *(__mmask64 *) __A = __builtin_ia32_kmovq (__B); | |
111 | } | |
112 | ||
6901ea62 AS |
113 | extern __inline __mmask32 |
114 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
115 | _knot_mask32 (__mmask32 __A) | |
116 | { | |
117 | return (__mmask32) __builtin_ia32_knotsi ((__mmask32) __A); | |
118 | } | |
119 | ||
120 | extern __inline __mmask64 | |
121 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
122 | _knot_mask64 (__mmask64 __A) | |
123 | { | |
124 | return (__mmask64) __builtin_ia32_knotdi ((__mmask64) __A); | |
125 | } | |
126 | ||
127 | extern __inline __mmask32 | |
128 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
129 | _kor_mask32 (__mmask32 __A, __mmask32 __B) | |
130 | { | |
131 | return (__mmask32) __builtin_ia32_korsi ((__mmask32) __A, (__mmask32) __B); | |
132 | } | |
133 | ||
134 | extern __inline __mmask64 | |
135 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
136 | _kor_mask64 (__mmask64 __A, __mmask64 __B) | |
137 | { | |
138 | return (__mmask64) __builtin_ia32_kordi ((__mmask64) __A, (__mmask64) __B); | |
139 | } | |
140 | ||
141 | extern __inline __mmask32 | |
142 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
143 | _kxnor_mask32 (__mmask32 __A, __mmask32 __B) | |
144 | { | |
145 | return (__mmask32) __builtin_ia32_kxnorsi ((__mmask32) __A, (__mmask32) __B); | |
146 | } | |
147 | ||
148 | extern __inline __mmask64 | |
149 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
150 | _kxnor_mask64 (__mmask64 __A, __mmask64 __B) | |
151 | { | |
152 | return (__mmask64) __builtin_ia32_kxnordi ((__mmask64) __A, (__mmask64) __B); | |
153 | } | |
154 | ||
155 | extern __inline __mmask32 | |
156 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
157 | _kxor_mask32 (__mmask32 __A, __mmask32 __B) | |
158 | { | |
159 | return (__mmask32) __builtin_ia32_kxorsi ((__mmask32) __A, (__mmask32) __B); | |
160 | } | |
161 | ||
162 | extern __inline __mmask64 | |
163 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
164 | _kxor_mask64 (__mmask64 __A, __mmask64 __B) | |
165 | { | |
166 | return (__mmask64) __builtin_ia32_kxordi ((__mmask64) __A, (__mmask64) __B); | |
167 | } | |
168 | ||
169 | extern __inline __mmask32 | |
170 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
171 | _kand_mask32 (__mmask32 __A, __mmask32 __B) | |
172 | { | |
173 | return (__mmask32) __builtin_ia32_kandsi ((__mmask32) __A, (__mmask32) __B); | |
174 | } | |
175 | ||
176 | extern __inline __mmask64 | |
177 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
178 | _kand_mask64 (__mmask64 __A, __mmask64 __B) | |
179 | { | |
180 | return (__mmask64) __builtin_ia32_kanddi ((__mmask64) __A, (__mmask64) __B); | |
181 | } | |
182 | ||
183 | extern __inline __mmask32 | |
184 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
185 | _kandn_mask32 (__mmask32 __A, __mmask32 __B) | |
186 | { | |
187 | return (__mmask32) __builtin_ia32_kandnsi ((__mmask32) __A, (__mmask32) __B); | |
188 | } | |
189 | ||
190 | extern __inline __mmask64 | |
191 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
192 | _kandn_mask64 (__mmask64 __A, __mmask64 __B) | |
193 | { | |
194 | return (__mmask64) __builtin_ia32_kandndi ((__mmask64) __A, (__mmask64) __B); | |
195 | } | |
196 | ||
936c0fe4 AI |
197 | extern __inline __m512i |
198 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
199 | _mm512_mask_mov_epi16 (__m512i __W, __mmask32 __U, __m512i __A) | |
200 | { | |
201 | return (__m512i) __builtin_ia32_movdquhi512_mask ((__v32hi) __A, | |
202 | (__v32hi) __W, | |
203 | (__mmask32) __U); | |
204 | } | |
205 | ||
206 | extern __inline __m512i | |
207 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
208 | _mm512_maskz_mov_epi16 (__mmask32 __U, __m512i __A) | |
209 | { | |
210 | return (__m512i) __builtin_ia32_movdquhi512_mask ((__v32hi) __A, | |
211 | (__v32hi) | |
f0515366 | 212 | _mm512_setzero_si512 (), |
936c0fe4 AI |
213 | (__mmask32) __U); |
214 | } | |
215 | ||
216 | extern __inline __m512i | |
217 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
218 | _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P) | |
219 | { | |
fc9cf6da | 220 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const short *) __P, |
936c0fe4 AI |
221 | (__v32hi) __W, |
222 | (__mmask32) __U); | |
223 | } | |
224 | ||
225 | extern __inline __m512i | |
226 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
227 | _mm512_maskz_loadu_epi16 (__mmask32 __U, void const *__P) | |
228 | { | |
fc9cf6da | 229 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const short *) __P, |
936c0fe4 | 230 | (__v32hi) |
f0515366 | 231 | _mm512_setzero_si512 (), |
936c0fe4 AI |
232 | (__mmask32) __U); |
233 | } | |
234 | ||
235 | extern __inline void | |
236 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
237 | _mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A) | |
238 | { | |
fc9cf6da | 239 | __builtin_ia32_storedquhi512_mask ((short *) __P, |
936c0fe4 AI |
240 | (__v32hi) __A, |
241 | (__mmask32) __U); | |
242 | } | |
243 | ||
244 | extern __inline __m512i | |
245 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
246 | _mm512_mask_mov_epi8 (__m512i __W, __mmask64 __U, __m512i __A) | |
247 | { | |
248 | return (__m512i) __builtin_ia32_movdquqi512_mask ((__v64qi) __A, | |
249 | (__v64qi) __W, | |
250 | (__mmask64) __U); | |
251 | } | |
252 | ||
253 | extern __inline __m512i | |
254 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
255 | _mm512_maskz_mov_epi8 (__mmask64 __U, __m512i __A) | |
256 | { | |
257 | return (__m512i) __builtin_ia32_movdquqi512_mask ((__v64qi) __A, | |
258 | (__v64qi) | |
f0515366 | 259 | _mm512_setzero_si512 (), |
936c0fe4 AI |
260 | (__mmask64) __U); |
261 | } | |
262 | ||
263 | extern __inline __mmask32 | |
264 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
265 | _mm512_kunpackw (__mmask32 __A, __mmask32 __B) | |
266 | { | |
267 | return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, | |
268 | (__mmask32) __B); | |
269 | } | |
270 | ||
6901ea62 AS |
271 | extern __inline __mmask32 |
272 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
273 | _kunpackw_mask32 (__mmask16 __A, __mmask16 __B) | |
274 | { | |
275 | return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, | |
276 | (__mmask32) __B); | |
277 | } | |
278 | ||
936c0fe4 AI |
279 | extern __inline __mmask64 |
280 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
281 | _mm512_kunpackd (__mmask64 __A, __mmask64 __B) | |
282 | { | |
283 | return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, | |
284 | (__mmask64) __B); | |
285 | } | |
286 | ||
6901ea62 AS |
287 | extern __inline __mmask64 |
288 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
289 | _kunpackd_mask64 (__mmask32 __A, __mmask32 __B) | |
290 | { | |
291 | return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, | |
292 | (__mmask64) __B); | |
293 | } | |
294 | ||
936c0fe4 AI |
295 | extern __inline __m512i |
296 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
297 | _mm512_mask_loadu_epi8 (__m512i __W, __mmask64 __U, void const *__P) | |
298 | { | |
fc9cf6da | 299 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const char *) __P, |
936c0fe4 AI |
300 | (__v64qi) __W, |
301 | (__mmask64) __U); | |
302 | } | |
303 | ||
304 | extern __inline __m512i | |
305 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
306 | _mm512_maskz_loadu_epi8 (__mmask64 __U, void const *__P) | |
307 | { | |
fc9cf6da | 308 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const char *) __P, |
936c0fe4 | 309 | (__v64qi) |
f0515366 | 310 | _mm512_setzero_si512 (), |
936c0fe4 AI |
311 | (__mmask64) __U); |
312 | } | |
313 | ||
314 | extern __inline void | |
315 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
316 | _mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A) | |
317 | { | |
fc9cf6da | 318 | __builtin_ia32_storedquqi512_mask ((char *) __P, |
936c0fe4 AI |
319 | (__v64qi) __A, |
320 | (__mmask64) __U); | |
321 | } | |
322 | ||
323 | extern __inline __m512i | |
324 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
325 | _mm512_sad_epu8 (__m512i __A, __m512i __B) | |
326 | { | |
327 | return (__m512i) __builtin_ia32_psadbw512 ((__v64qi) __A, | |
328 | (__v64qi) __B); | |
329 | } | |
330 | ||
331 | extern __inline __m256i | |
332 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
333 | _mm512_cvtepi16_epi8 (__m512i __A) | |
334 | { | |
335 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, | |
336 | (__v32qi) _mm256_undefined_si256(), | |
337 | (__mmask32) -1); | |
338 | } | |
339 | ||
340 | extern __inline __m256i | |
341 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
342 | _mm512_mask_cvtepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) | |
343 | { | |
344 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, | |
345 | (__v32qi) __O, __M); | |
346 | } | |
347 | ||
348 | extern __inline __m256i | |
349 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
350 | _mm512_maskz_cvtepi16_epi8 (__mmask32 __M, __m512i __A) | |
351 | { | |
352 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, | |
353 | (__v32qi) | |
354 | _mm256_setzero_si256 (), | |
355 | __M); | |
356 | } | |
357 | ||
358 | extern __inline __m256i | |
359 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
360 | _mm512_cvtsepi16_epi8 (__m512i __A) | |
361 | { | |
362 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, | |
363 | (__v32qi)_mm256_undefined_si256(), | |
364 | (__mmask32) -1); | |
365 | } | |
366 | ||
367 | extern __inline __m256i | |
368 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
369 | _mm512_mask_cvtsepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) | |
370 | { | |
371 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, | |
372 | (__v32qi)__O, | |
373 | __M); | |
374 | } | |
375 | ||
376 | extern __inline __m256i | |
377 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
378 | _mm512_maskz_cvtsepi16_epi8 (__mmask32 __M, __m512i __A) | |
379 | { | |
380 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, | |
381 | (__v32qi) | |
382 | _mm256_setzero_si256 (), | |
383 | __M); | |
384 | } | |
385 | ||
386 | extern __inline __m256i | |
387 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
388 | _mm512_cvtusepi16_epi8 (__m512i __A) | |
389 | { | |
390 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, | |
391 | (__v32qi)_mm256_undefined_si256(), | |
392 | (__mmask32) -1); | |
393 | } | |
394 | ||
395 | extern __inline __m256i | |
396 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
397 | _mm512_mask_cvtusepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) | |
398 | { | |
399 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, | |
400 | (__v32qi) __O, | |
401 | __M); | |
402 | } | |
403 | ||
404 | extern __inline __m256i | |
405 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
406 | _mm512_maskz_cvtusepi16_epi8 (__mmask32 __M, __m512i __A) | |
407 | { | |
408 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, | |
409 | (__v32qi) | |
410 | _mm256_setzero_si256 (), | |
411 | __M); | |
412 | } | |
413 | ||
414 | extern __inline __m512i | |
415 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
416 | _mm512_broadcastb_epi8 (__m128i __A) | |
417 | { | |
418 | return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, | |
4271e5cb UB |
419 | (__v64qi)_mm512_undefined_epi32(), |
420 | (__mmask64) -1); | |
936c0fe4 AI |
421 | } |
422 | ||
423 | extern __inline __m512i | |
424 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
425 | _mm512_mask_broadcastb_epi8 (__m512i __O, __mmask64 __M, __m128i __A) | |
426 | { | |
427 | return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, | |
428 | (__v64qi) __O, | |
429 | __M); | |
430 | } | |
431 | ||
432 | extern __inline __m512i | |
433 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
434 | _mm512_maskz_broadcastb_epi8 (__mmask64 __M, __m128i __A) | |
435 | { | |
436 | return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, | |
437 | (__v64qi) | |
f0515366 | 438 | _mm512_setzero_si512 (), |
936c0fe4 AI |
439 | __M); |
440 | } | |
441 | ||
442 | extern __inline __m512i | |
443 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
444 | _mm512_mask_set1_epi8 (__m512i __O, __mmask64 __M, char __A) | |
445 | { | |
446 | return (__m512i) __builtin_ia32_pbroadcastb512_gpr_mask (__A, | |
447 | (__v64qi) __O, | |
448 | __M); | |
449 | } | |
450 | ||
451 | extern __inline __m512i | |
452 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
453 | _mm512_maskz_set1_epi8 (__mmask64 __M, char __A) | |
454 | { | |
f0515366 JJ |
455 | return (__m512i) |
456 | __builtin_ia32_pbroadcastb512_gpr_mask (__A, | |
457 | (__v64qi) | |
458 | _mm512_setzero_si512 (), | |
459 | __M); | |
936c0fe4 AI |
460 | } |
461 | ||
462 | extern __inline __m512i | |
463 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
464 | _mm512_broadcastw_epi16 (__m128i __A) | |
465 | { | |
466 | return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, | |
4271e5cb UB |
467 | (__v32hi)_mm512_undefined_epi32(), |
468 | (__mmask32) -1); | |
936c0fe4 AI |
469 | } |
470 | ||
471 | extern __inline __m512i | |
472 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
473 | _mm512_mask_broadcastw_epi16 (__m512i __O, __mmask32 __M, __m128i __A) | |
474 | { | |
475 | return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, | |
476 | (__v32hi) __O, | |
477 | __M); | |
478 | } | |
479 | ||
480 | extern __inline __m512i | |
481 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
482 | _mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A) | |
483 | { | |
484 | return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, | |
485 | (__v32hi) | |
f0515366 | 486 | _mm512_setzero_si512 (), |
936c0fe4 AI |
487 | __M); |
488 | } | |
489 | ||
490 | extern __inline __m512i | |
491 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
492 | _mm512_mask_set1_epi16 (__m512i __O, __mmask32 __M, short __A) | |
493 | { | |
494 | return (__m512i) __builtin_ia32_pbroadcastw512_gpr_mask (__A, | |
495 | (__v32hi) __O, | |
496 | __M); | |
497 | } | |
498 | ||
499 | extern __inline __m512i | |
500 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
501 | _mm512_maskz_set1_epi16 (__mmask32 __M, short __A) | |
502 | { | |
f0515366 JJ |
503 | return (__m512i) |
504 | __builtin_ia32_pbroadcastw512_gpr_mask (__A, | |
505 | (__v32hi) | |
506 | _mm512_setzero_si512 (), | |
507 | __M); | |
936c0fe4 AI |
508 | } |
509 | ||
510 | extern __inline __m512i | |
511 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
512 | _mm512_mulhrs_epi16 (__m512i __A, __m512i __B) | |
513 | { | |
514 | return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, | |
515 | (__v32hi) __B, | |
516 | (__v32hi) | |
f0515366 | 517 | _mm512_setzero_si512 (), |
936c0fe4 AI |
518 | (__mmask32) -1); |
519 | } | |
520 | ||
521 | extern __inline __m512i | |
522 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
523 | _mm512_mask_mulhrs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
524 | __m512i __B) | |
525 | { | |
526 | return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, | |
527 | (__v32hi) __B, | |
528 | (__v32hi) __W, | |
529 | (__mmask32) __U); | |
530 | } | |
531 | ||
532 | extern __inline __m512i | |
533 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
534 | _mm512_maskz_mulhrs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
535 | { | |
536 | return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, | |
537 | (__v32hi) __B, | |
538 | (__v32hi) | |
f0515366 | 539 | _mm512_setzero_si512 (), |
936c0fe4 AI |
540 | (__mmask32) __U); |
541 | } | |
542 | ||
543 | extern __inline __m512i | |
544 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
545 | _mm512_mulhi_epi16 (__m512i __A, __m512i __B) | |
546 | { | |
547 | return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, | |
548 | (__v32hi) __B, | |
549 | (__v32hi) | |
f0515366 | 550 | _mm512_setzero_si512 (), |
936c0fe4 AI |
551 | (__mmask32) -1); |
552 | } | |
553 | ||
554 | extern __inline __m512i | |
555 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
556 | _mm512_mask_mulhi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
557 | __m512i __B) | |
558 | { | |
559 | return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, | |
560 | (__v32hi) __B, | |
561 | (__v32hi) __W, | |
562 | (__mmask32) __U); | |
563 | } | |
564 | ||
565 | extern __inline __m512i | |
566 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
567 | _mm512_maskz_mulhi_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
568 | { | |
569 | return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, | |
570 | (__v32hi) __B, | |
571 | (__v32hi) | |
f0515366 | 572 | _mm512_setzero_si512 (), |
936c0fe4 AI |
573 | (__mmask32) __U); |
574 | } | |
575 | ||
576 | extern __inline __m512i | |
577 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
578 | _mm512_mulhi_epu16 (__m512i __A, __m512i __B) | |
579 | { | |
580 | return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, | |
581 | (__v32hi) __B, | |
582 | (__v32hi) | |
f0515366 | 583 | _mm512_setzero_si512 (), |
936c0fe4 AI |
584 | (__mmask32) -1); |
585 | } | |
586 | ||
587 | extern __inline __m512i | |
588 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
589 | _mm512_mask_mulhi_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
590 | __m512i __B) | |
591 | { | |
592 | return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, | |
593 | (__v32hi) __B, | |
594 | (__v32hi) __W, | |
595 | (__mmask32) __U); | |
596 | } | |
597 | ||
598 | extern __inline __m512i | |
599 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
600 | _mm512_maskz_mulhi_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
601 | { | |
602 | return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, | |
603 | (__v32hi) __B, | |
604 | (__v32hi) | |
f0515366 | 605 | _mm512_setzero_si512 (), |
936c0fe4 AI |
606 | (__mmask32) __U); |
607 | } | |
608 | ||
609 | extern __inline __m512i | |
610 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
611 | _mm512_mullo_epi16 (__m512i __A, __m512i __B) | |
612 | { | |
2069d6fc | 613 | return (__m512i) ((__v32hu) __A * (__v32hu) __B); |
936c0fe4 AI |
614 | } |
615 | ||
616 | extern __inline __m512i | |
617 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
618 | _mm512_mask_mullo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
619 | __m512i __B) | |
620 | { | |
621 | return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A, | |
622 | (__v32hi) __B, | |
623 | (__v32hi) __W, | |
624 | (__mmask32) __U); | |
625 | } | |
626 | ||
627 | extern __inline __m512i | |
628 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
629 | _mm512_maskz_mullo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
630 | { | |
631 | return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A, | |
632 | (__v32hi) __B, | |
633 | (__v32hi) | |
f0515366 | 634 | _mm512_setzero_si512 (), |
936c0fe4 AI |
635 | (__mmask32) __U); |
636 | } | |
637 | ||
638 | extern __inline __m512i | |
639 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
640 | _mm512_cvtepi8_epi16 (__m256i __A) | |
641 | { | |
642 | return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, | |
643 | (__v32hi) | |
f0515366 | 644 | _mm512_setzero_si512 (), |
936c0fe4 AI |
645 | (__mmask32) -1); |
646 | } | |
647 | ||
648 | extern __inline __m512i | |
649 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
650 | _mm512_mask_cvtepi8_epi16 (__m512i __W, __mmask32 __U, __m256i __A) | |
651 | { | |
652 | return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, | |
653 | (__v32hi) __W, | |
654 | (__mmask32) __U); | |
655 | } | |
656 | ||
657 | extern __inline __m512i | |
658 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
659 | _mm512_maskz_cvtepi8_epi16 (__mmask32 __U, __m256i __A) | |
660 | { | |
661 | return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, | |
662 | (__v32hi) | |
f0515366 | 663 | _mm512_setzero_si512 (), |
936c0fe4 AI |
664 | (__mmask32) __U); |
665 | } | |
666 | ||
667 | extern __inline __m512i | |
668 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
669 | _mm512_cvtepu8_epi16 (__m256i __A) | |
670 | { | |
671 | return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, | |
672 | (__v32hi) | |
f0515366 | 673 | _mm512_setzero_si512 (), |
936c0fe4 AI |
674 | (__mmask32) -1); |
675 | } | |
676 | ||
677 | extern __inline __m512i | |
678 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
679 | _mm512_mask_cvtepu8_epi16 (__m512i __W, __mmask32 __U, __m256i __A) | |
680 | { | |
681 | return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, | |
682 | (__v32hi) __W, | |
683 | (__mmask32) __U); | |
684 | } | |
685 | ||
686 | extern __inline __m512i | |
687 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
688 | _mm512_maskz_cvtepu8_epi16 (__mmask32 __U, __m256i __A) | |
689 | { | |
690 | return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, | |
691 | (__v32hi) | |
f0515366 | 692 | _mm512_setzero_si512 (), |
936c0fe4 AI |
693 | (__mmask32) __U); |
694 | } | |
695 | ||
696 | extern __inline __m512i | |
697 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
698 | _mm512_permutexvar_epi16 (__m512i __A, __m512i __B) | |
699 | { | |
700 | return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, | |
701 | (__v32hi) __A, | |
702 | (__v32hi) | |
f0515366 | 703 | _mm512_setzero_si512 (), |
936c0fe4 AI |
704 | (__mmask32) -1); |
705 | } | |
706 | ||
707 | extern __inline __m512i | |
708 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
709 | _mm512_maskz_permutexvar_epi16 (__mmask32 __M, __m512i __A, | |
710 | __m512i __B) | |
711 | { | |
712 | return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, | |
713 | (__v32hi) __A, | |
714 | (__v32hi) | |
f0515366 | 715 | _mm512_setzero_si512 (), |
936c0fe4 AI |
716 | (__mmask32) __M); |
717 | } | |
718 | ||
719 | extern __inline __m512i | |
720 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
721 | _mm512_mask_permutexvar_epi16 (__m512i __W, __mmask32 __M, __m512i __A, | |
722 | __m512i __B) | |
723 | { | |
724 | return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, | |
725 | (__v32hi) __A, | |
726 | (__v32hi) __W, | |
727 | (__mmask32) __M); | |
728 | } | |
729 | ||
730 | extern __inline __m512i | |
731 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
732 | _mm512_permutex2var_epi16 (__m512i __A, __m512i __I, __m512i __B) | |
733 | { | |
734 | return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I | |
735 | /* idx */ , | |
736 | (__v32hi) __A, | |
737 | (__v32hi) __B, | |
281f7120 | 738 | (__mmask32) -1); |
936c0fe4 AI |
739 | } |
740 | ||
741 | extern __inline __m512i | |
742 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
743 | _mm512_mask_permutex2var_epi16 (__m512i __A, __mmask32 __U, | |
744 | __m512i __I, __m512i __B) | |
745 | { | |
746 | return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I | |
747 | /* idx */ , | |
748 | (__v32hi) __A, | |
749 | (__v32hi) __B, | |
750 | (__mmask32) | |
751 | __U); | |
752 | } | |
753 | ||
754 | extern __inline __m512i | |
755 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
756 | _mm512_mask2_permutex2var_epi16 (__m512i __A, __m512i __I, | |
757 | __mmask32 __U, __m512i __B) | |
758 | { | |
759 | return (__m512i) __builtin_ia32_vpermi2varhi512_mask ((__v32hi) __A, | |
760 | (__v32hi) __I | |
761 | /* idx */ , | |
762 | (__v32hi) __B, | |
763 | (__mmask32) | |
764 | __U); | |
765 | } | |
766 | ||
767 | extern __inline __m512i | |
768 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
769 | _mm512_maskz_permutex2var_epi16 (__mmask32 __U, __m512i __A, | |
770 | __m512i __I, __m512i __B) | |
771 | { | |
772 | return (__m512i) __builtin_ia32_vpermt2varhi512_maskz ((__v32hi) __I | |
773 | /* idx */ , | |
774 | (__v32hi) __A, | |
775 | (__v32hi) __B, | |
776 | (__mmask32) | |
777 | __U); | |
778 | } | |
779 | ||
780 | extern __inline __m512i | |
781 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
782 | _mm512_avg_epu8 (__m512i __A, __m512i __B) | |
783 | { | |
784 | return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, | |
785 | (__v64qi) __B, | |
786 | (__v64qi) | |
f0515366 | 787 | _mm512_setzero_si512 (), |
936c0fe4 AI |
788 | (__mmask64) -1); |
789 | } | |
790 | ||
791 | extern __inline __m512i | |
792 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
793 | _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A, | |
794 | __m512i __B) | |
795 | { | |
796 | return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, | |
797 | (__v64qi) __B, | |
798 | (__v64qi) __W, | |
799 | (__mmask64) __U); | |
800 | } | |
801 | ||
802 | extern __inline __m512i | |
803 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
804 | _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B) | |
805 | { | |
806 | return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, | |
807 | (__v64qi) __B, | |
808 | (__v64qi) | |
f0515366 | 809 | _mm512_setzero_si512 (), |
936c0fe4 AI |
810 | (__mmask64) __U); |
811 | } | |
812 | ||
813 | extern __inline __m512i | |
814 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
815 | _mm512_add_epi8 (__m512i __A, __m512i __B) | |
816 | { | |
2069d6fc | 817 | return (__m512i) ((__v64qu) __A + (__v64qu) __B); |
936c0fe4 AI |
818 | } |
819 | ||
820 | extern __inline __m512i | |
821 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
822 | _mm512_mask_add_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
823 | __m512i __B) | |
824 | { | |
825 | return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A, | |
826 | (__v64qi) __B, | |
827 | (__v64qi) __W, | |
828 | (__mmask64) __U); | |
829 | } | |
830 | ||
831 | extern __inline __m512i | |
832 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
833 | _mm512_maskz_add_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
834 | { | |
835 | return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A, | |
836 | (__v64qi) __B, | |
837 | (__v64qi) | |
f0515366 | 838 | _mm512_setzero_si512 (), |
936c0fe4 AI |
839 | (__mmask64) __U); |
840 | } | |
841 | ||
842 | extern __inline __m512i | |
843 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
844 | _mm512_sub_epi8 (__m512i __A, __m512i __B) | |
845 | { | |
2069d6fc | 846 | return (__m512i) ((__v64qu) __A - (__v64qu) __B); |
936c0fe4 AI |
847 | } |
848 | ||
849 | extern __inline __m512i | |
850 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
851 | _mm512_mask_sub_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
852 | __m512i __B) | |
853 | { | |
854 | return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A, | |
855 | (__v64qi) __B, | |
856 | (__v64qi) __W, | |
857 | (__mmask64) __U); | |
858 | } | |
859 | ||
860 | extern __inline __m512i | |
861 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
862 | _mm512_maskz_sub_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
863 | { | |
864 | return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A, | |
865 | (__v64qi) __B, | |
866 | (__v64qi) | |
f0515366 | 867 | _mm512_setzero_si512 (), |
936c0fe4 AI |
868 | (__mmask64) __U); |
869 | } | |
870 | ||
871 | extern __inline __m512i | |
872 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
873 | _mm512_avg_epu16 (__m512i __A, __m512i __B) | |
874 | { | |
875 | return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, | |
876 | (__v32hi) __B, | |
877 | (__v32hi) | |
f0515366 | 878 | _mm512_setzero_si512 (), |
936c0fe4 AI |
879 | (__mmask32) -1); |
880 | } | |
881 | ||
882 | extern __inline __m512i | |
883 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
884 | _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
885 | __m512i __B) | |
886 | { | |
887 | return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, | |
888 | (__v32hi) __B, | |
889 | (__v32hi) __W, | |
890 | (__mmask32) __U); | |
891 | } | |
892 | ||
893 | extern __inline __m512i | |
894 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
895 | _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
896 | { | |
897 | return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, | |
898 | (__v32hi) __B, | |
899 | (__v32hi) | |
f0515366 | 900 | _mm512_setzero_si512 (), |
936c0fe4 AI |
901 | (__mmask32) __U); |
902 | } | |
903 | ||
904 | extern __inline __m512i | |
905 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
906 | _mm512_subs_epi8 (__m512i __A, __m512i __B) | |
907 | { | |
908 | return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, | |
909 | (__v64qi) __B, | |
910 | (__v64qi) | |
f0515366 | 911 | _mm512_setzero_si512 (), |
936c0fe4 AI |
912 | (__mmask64) -1); |
913 | } | |
914 | ||
915 | extern __inline __m512i | |
916 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
917 | _mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
918 | __m512i __B) | |
919 | { | |
920 | return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, | |
921 | (__v64qi) __B, | |
922 | (__v64qi) __W, | |
923 | (__mmask64) __U); | |
924 | } | |
925 | ||
926 | extern __inline __m512i | |
927 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
928 | _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
929 | { | |
930 | return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, | |
931 | (__v64qi) __B, | |
932 | (__v64qi) | |
f0515366 | 933 | _mm512_setzero_si512 (), |
936c0fe4 AI |
934 | (__mmask64) __U); |
935 | } | |
936 | ||
937 | extern __inline __m512i | |
938 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
939 | _mm512_subs_epu8 (__m512i __A, __m512i __B) | |
940 | { | |
941 | return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, | |
942 | (__v64qi) __B, | |
943 | (__v64qi) | |
f0515366 | 944 | _mm512_setzero_si512 (), |
936c0fe4 AI |
945 | (__mmask64) -1); |
946 | } | |
947 | ||
948 | extern __inline __m512i | |
949 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
950 | _mm512_mask_subs_epu8 (__m512i __W, __mmask64 __U, __m512i __A, | |
951 | __m512i __B) | |
952 | { | |
953 | return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, | |
954 | (__v64qi) __B, | |
955 | (__v64qi) __W, | |
956 | (__mmask64) __U); | |
957 | } | |
958 | ||
959 | extern __inline __m512i | |
960 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
961 | _mm512_maskz_subs_epu8 (__mmask64 __U, __m512i __A, __m512i __B) | |
962 | { | |
963 | return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, | |
964 | (__v64qi) __B, | |
965 | (__v64qi) | |
f0515366 | 966 | _mm512_setzero_si512 (), |
936c0fe4 AI |
967 | (__mmask64) __U); |
968 | } | |
969 | ||
970 | extern __inline __m512i | |
971 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
972 | _mm512_adds_epi8 (__m512i __A, __m512i __B) | |
973 | { | |
974 | return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, | |
975 | (__v64qi) __B, | |
976 | (__v64qi) | |
f0515366 | 977 | _mm512_setzero_si512 (), |
936c0fe4 AI |
978 | (__mmask64) -1); |
979 | } | |
980 | ||
981 | extern __inline __m512i | |
982 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
983 | _mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
984 | __m512i __B) | |
985 | { | |
986 | return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, | |
987 | (__v64qi) __B, | |
988 | (__v64qi) __W, | |
989 | (__mmask64) __U); | |
990 | } | |
991 | ||
992 | extern __inline __m512i | |
993 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
994 | _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
995 | { | |
996 | return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, | |
997 | (__v64qi) __B, | |
998 | (__v64qi) | |
f0515366 | 999 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1000 | (__mmask64) __U); |
1001 | } | |
1002 | ||
1003 | extern __inline __m512i | |
1004 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1005 | _mm512_adds_epu8 (__m512i __A, __m512i __B) | |
1006 | { | |
1007 | return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, | |
1008 | (__v64qi) __B, | |
1009 | (__v64qi) | |
f0515366 | 1010 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1011 | (__mmask64) -1); |
1012 | } | |
1013 | ||
1014 | extern __inline __m512i | |
1015 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1016 | _mm512_mask_adds_epu8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1017 | __m512i __B) | |
1018 | { | |
1019 | return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, | |
1020 | (__v64qi) __B, | |
1021 | (__v64qi) __W, | |
1022 | (__mmask64) __U); | |
1023 | } | |
1024 | ||
1025 | extern __inline __m512i | |
1026 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1027 | _mm512_maskz_adds_epu8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1028 | { | |
1029 | return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, | |
1030 | (__v64qi) __B, | |
1031 | (__v64qi) | |
f0515366 | 1032 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1033 | (__mmask64) __U); |
1034 | } | |
1035 | ||
1036 | extern __inline __m512i | |
1037 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1038 | _mm512_sub_epi16 (__m512i __A, __m512i __B) | |
1039 | { | |
2069d6fc | 1040 | return (__m512i) ((__v32hu) __A - (__v32hu) __B); |
936c0fe4 AI |
1041 | } |
1042 | ||
1043 | extern __inline __m512i | |
1044 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1045 | _mm512_mask_sub_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1046 | __m512i __B) | |
1047 | { | |
1048 | return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A, | |
1049 | (__v32hi) __B, | |
1050 | (__v32hi) __W, | |
1051 | (__mmask32) __U); | |
1052 | } | |
1053 | ||
1054 | extern __inline __m512i | |
1055 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1056 | _mm512_maskz_sub_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1057 | { | |
1058 | return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A, | |
1059 | (__v32hi) __B, | |
1060 | (__v32hi) | |
f0515366 | 1061 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1062 | (__mmask32) __U); |
1063 | } | |
1064 | ||
1065 | extern __inline __m512i | |
1066 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1067 | _mm512_subs_epi16 (__m512i __A, __m512i __B) | |
1068 | { | |
1069 | return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, | |
1070 | (__v32hi) __B, | |
1071 | (__v32hi) | |
f0515366 | 1072 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1073 | (__mmask32) -1); |
1074 | } | |
1075 | ||
1076 | extern __inline __m512i | |
1077 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1078 | _mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1079 | __m512i __B) | |
1080 | { | |
1081 | return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, | |
1082 | (__v32hi) __B, | |
1083 | (__v32hi) __W, | |
1084 | (__mmask32) __U); | |
1085 | } | |
1086 | ||
1087 | extern __inline __m512i | |
1088 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1089 | _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1090 | { | |
1091 | return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, | |
1092 | (__v32hi) __B, | |
1093 | (__v32hi) | |
f0515366 | 1094 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1095 | (__mmask32) __U); |
1096 | } | |
1097 | ||
1098 | extern __inline __m512i | |
1099 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1100 | _mm512_subs_epu16 (__m512i __A, __m512i __B) | |
1101 | { | |
1102 | return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, | |
1103 | (__v32hi) __B, | |
1104 | (__v32hi) | |
f0515366 | 1105 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1106 | (__mmask32) -1); |
1107 | } | |
1108 | ||
1109 | extern __inline __m512i | |
1110 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1111 | _mm512_mask_subs_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1112 | __m512i __B) | |
1113 | { | |
1114 | return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, | |
1115 | (__v32hi) __B, | |
1116 | (__v32hi) __W, | |
1117 | (__mmask32) __U); | |
1118 | } | |
1119 | ||
1120 | extern __inline __m512i | |
1121 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1122 | _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1123 | { | |
1124 | return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, | |
1125 | (__v32hi) __B, | |
1126 | (__v32hi) | |
f0515366 | 1127 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1128 | (__mmask32) __U); |
1129 | } | |
1130 | ||
1131 | extern __inline __m512i | |
1132 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1133 | _mm512_add_epi16 (__m512i __A, __m512i __B) | |
1134 | { | |
2069d6fc | 1135 | return (__m512i) ((__v32hu) __A + (__v32hu) __B); |
936c0fe4 AI |
1136 | } |
1137 | ||
1138 | extern __inline __m512i | |
1139 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1140 | _mm512_mask_add_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1141 | __m512i __B) | |
1142 | { | |
1143 | return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A, | |
1144 | (__v32hi) __B, | |
1145 | (__v32hi) __W, | |
1146 | (__mmask32) __U); | |
1147 | } | |
1148 | ||
1149 | extern __inline __m512i | |
1150 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1151 | _mm512_maskz_add_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1152 | { | |
1153 | return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A, | |
1154 | (__v32hi) __B, | |
1155 | (__v32hi) | |
f0515366 | 1156 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1157 | (__mmask32) __U); |
1158 | } | |
1159 | ||
1160 | extern __inline __m512i | |
1161 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1162 | _mm512_adds_epi16 (__m512i __A, __m512i __B) | |
1163 | { | |
1164 | return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, | |
1165 | (__v32hi) __B, | |
1166 | (__v32hi) | |
f0515366 | 1167 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1168 | (__mmask32) -1); |
1169 | } | |
1170 | ||
1171 | extern __inline __m512i | |
1172 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1173 | _mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1174 | __m512i __B) | |
1175 | { | |
1176 | return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, | |
1177 | (__v32hi) __B, | |
1178 | (__v32hi) __W, | |
1179 | (__mmask32) __U); | |
1180 | } | |
1181 | ||
1182 | extern __inline __m512i | |
1183 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1184 | _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1185 | { | |
1186 | return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, | |
1187 | (__v32hi) __B, | |
1188 | (__v32hi) | |
f0515366 | 1189 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1190 | (__mmask32) __U); |
1191 | } | |
1192 | ||
1193 | extern __inline __m512i | |
1194 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1195 | _mm512_adds_epu16 (__m512i __A, __m512i __B) | |
1196 | { | |
1197 | return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, | |
1198 | (__v32hi) __B, | |
1199 | (__v32hi) | |
f0515366 | 1200 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1201 | (__mmask32) -1); |
1202 | } | |
1203 | ||
1204 | extern __inline __m512i | |
1205 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1206 | _mm512_mask_adds_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1207 | __m512i __B) | |
1208 | { | |
1209 | return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, | |
1210 | (__v32hi) __B, | |
1211 | (__v32hi) __W, | |
1212 | (__mmask32) __U); | |
1213 | } | |
1214 | ||
1215 | extern __inline __m512i | |
1216 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1217 | _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1218 | { | |
1219 | return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, | |
1220 | (__v32hi) __B, | |
1221 | (__v32hi) | |
f0515366 | 1222 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1223 | (__mmask32) __U); |
1224 | } | |
1225 | ||
1226 | extern __inline __m512i | |
1227 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1228 | _mm512_srl_epi16 (__m512i __A, __m128i __B) | |
1229 | { | |
1230 | return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, | |
1231 | (__v8hi) __B, | |
1232 | (__v32hi) | |
f0515366 | 1233 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1234 | (__mmask32) -1); |
1235 | } | |
1236 | ||
1237 | extern __inline __m512i | |
1238 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1239 | _mm512_mask_srl_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1240 | __m128i __B) | |
1241 | { | |
1242 | return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, | |
1243 | (__v8hi) __B, | |
1244 | (__v32hi) __W, | |
1245 | (__mmask32) __U); | |
1246 | } | |
1247 | ||
1248 | extern __inline __m512i | |
1249 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1250 | _mm512_maskz_srl_epi16 (__mmask32 __U, __m512i __A, __m128i __B) | |
1251 | { | |
1252 | return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, | |
1253 | (__v8hi) __B, | |
1254 | (__v32hi) | |
f0515366 | 1255 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1256 | (__mmask32) __U); |
1257 | } | |
1258 | ||
1259 | extern __inline __m512i | |
1260 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1261 | _mm512_packs_epi16 (__m512i __A, __m512i __B) | |
1262 | { | |
1263 | return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, | |
1264 | (__v32hi) __B, | |
1265 | (__v64qi) | |
f0515366 | 1266 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1267 | (__mmask64) -1); |
1268 | } | |
1269 | ||
1270 | extern __inline __m512i | |
1271 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1272 | _mm512_sll_epi16 (__m512i __A, __m128i __B) | |
1273 | { | |
1274 | return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, | |
1275 | (__v8hi) __B, | |
1276 | (__v32hi) | |
f0515366 | 1277 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1278 | (__mmask32) -1); |
1279 | } | |
1280 | ||
1281 | extern __inline __m512i | |
1282 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1283 | _mm512_mask_sll_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1284 | __m128i __B) | |
1285 | { | |
1286 | return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, | |
1287 | (__v8hi) __B, | |
1288 | (__v32hi) __W, | |
1289 | (__mmask32) __U); | |
1290 | } | |
1291 | ||
1292 | extern __inline __m512i | |
1293 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1294 | _mm512_maskz_sll_epi16 (__mmask32 __U, __m512i __A, __m128i __B) | |
1295 | { | |
1296 | return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, | |
1297 | (__v8hi) __B, | |
1298 | (__v32hi) | |
f0515366 | 1299 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1300 | (__mmask32) __U); |
1301 | } | |
1302 | ||
1303 | extern __inline __m512i | |
1304 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1305 | _mm512_maddubs_epi16 (__m512i __X, __m512i __Y) | |
1306 | { | |
1307 | return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, | |
1308 | (__v64qi) __Y, | |
1309 | (__v32hi) | |
f0515366 | 1310 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1311 | (__mmask32) -1); |
1312 | } | |
1313 | ||
1314 | extern __inline __m512i | |
1315 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1316 | _mm512_mask_maddubs_epi16 (__m512i __W, __mmask32 __U, __m512i __X, | |
1317 | __m512i __Y) | |
1318 | { | |
1319 | return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, | |
1320 | (__v64qi) __Y, | |
1321 | (__v32hi) __W, | |
1322 | (__mmask32) __U); | |
1323 | } | |
1324 | ||
1325 | extern __inline __m512i | |
1326 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1327 | _mm512_maskz_maddubs_epi16 (__mmask32 __U, __m512i __X, __m512i __Y) | |
1328 | { | |
1329 | return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, | |
1330 | (__v64qi) __Y, | |
1331 | (__v32hi) | |
f0515366 | 1332 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1333 | (__mmask32) __U); |
1334 | } | |
1335 | ||
1336 | extern __inline __m512i | |
1337 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1338 | _mm512_madd_epi16 (__m512i __A, __m512i __B) | |
1339 | { | |
1340 | return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, | |
1341 | (__v32hi) __B, | |
1342 | (__v16si) | |
1343 | _mm512_setzero_si512 (), | |
1344 | (__mmask16) -1); | |
1345 | } | |
1346 | ||
1347 | extern __inline __m512i | |
1348 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1349 | _mm512_mask_madd_epi16 (__m512i __W, __mmask16 __U, __m512i __A, | |
1350 | __m512i __B) | |
1351 | { | |
1352 | return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, | |
1353 | (__v32hi) __B, | |
1354 | (__v16si) __W, | |
1355 | (__mmask16) __U); | |
1356 | } | |
1357 | ||
1358 | extern __inline __m512i | |
1359 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1360 | _mm512_maskz_madd_epi16 (__mmask16 __U, __m512i __A, __m512i __B) | |
1361 | { | |
1362 | return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, | |
1363 | (__v32hi) __B, | |
1364 | (__v16si) | |
1365 | _mm512_setzero_si512 (), | |
1366 | (__mmask16) __U); | |
1367 | } | |
1368 | ||
1369 | extern __inline __m512i | |
1370 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1371 | _mm512_unpackhi_epi8 (__m512i __A, __m512i __B) | |
1372 | { | |
1373 | return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, | |
1374 | (__v64qi) __B, | |
1375 | (__v64qi) | |
f0515366 | 1376 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1377 | (__mmask64) -1); |
1378 | } | |
1379 | ||
1380 | extern __inline __m512i | |
1381 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1382 | _mm512_mask_unpackhi_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1383 | __m512i __B) | |
1384 | { | |
1385 | return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, | |
1386 | (__v64qi) __B, | |
1387 | (__v64qi) __W, | |
1388 | (__mmask64) __U); | |
1389 | } | |
1390 | ||
1391 | extern __inline __m512i | |
1392 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1393 | _mm512_maskz_unpackhi_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1394 | { | |
1395 | return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, | |
1396 | (__v64qi) __B, | |
1397 | (__v64qi) | |
f0515366 | 1398 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1399 | (__mmask64) __U); |
1400 | } | |
1401 | ||
1402 | extern __inline __m512i | |
1403 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1404 | _mm512_unpackhi_epi16 (__m512i __A, __m512i __B) | |
1405 | { | |
1406 | return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, | |
1407 | (__v32hi) __B, | |
1408 | (__v32hi) | |
f0515366 | 1409 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1410 | (__mmask32) -1); |
1411 | } | |
1412 | ||
1413 | extern __inline __m512i | |
1414 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1415 | _mm512_mask_unpackhi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1416 | __m512i __B) | |
1417 | { | |
1418 | return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, | |
1419 | (__v32hi) __B, | |
1420 | (__v32hi) __W, | |
1421 | (__mmask32) __U); | |
1422 | } | |
1423 | ||
1424 | extern __inline __m512i | |
1425 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1426 | _mm512_maskz_unpackhi_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1427 | { | |
1428 | return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, | |
1429 | (__v32hi) __B, | |
1430 | (__v32hi) | |
f0515366 | 1431 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1432 | (__mmask32) __U); |
1433 | } | |
1434 | ||
1435 | extern __inline __m512i | |
1436 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1437 | _mm512_unpacklo_epi8 (__m512i __A, __m512i __B) | |
1438 | { | |
1439 | return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, | |
1440 | (__v64qi) __B, | |
1441 | (__v64qi) | |
f0515366 | 1442 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1443 | (__mmask64) -1); |
1444 | } | |
1445 | ||
1446 | extern __inline __m512i | |
1447 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1448 | _mm512_mask_unpacklo_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1449 | __m512i __B) | |
1450 | { | |
1451 | return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, | |
1452 | (__v64qi) __B, | |
1453 | (__v64qi) __W, | |
1454 | (__mmask64) __U); | |
1455 | } | |
1456 | ||
1457 | extern __inline __m512i | |
1458 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1459 | _mm512_maskz_unpacklo_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1460 | { | |
1461 | return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, | |
1462 | (__v64qi) __B, | |
1463 | (__v64qi) | |
f0515366 | 1464 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1465 | (__mmask64) __U); |
1466 | } | |
1467 | ||
1468 | extern __inline __m512i | |
1469 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1470 | _mm512_unpacklo_epi16 (__m512i __A, __m512i __B) | |
1471 | { | |
1472 | return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, | |
1473 | (__v32hi) __B, | |
1474 | (__v32hi) | |
f0515366 | 1475 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1476 | (__mmask32) -1); |
1477 | } | |
1478 | ||
1479 | extern __inline __m512i | |
1480 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1481 | _mm512_mask_unpacklo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1482 | __m512i __B) | |
1483 | { | |
1484 | return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, | |
1485 | (__v32hi) __B, | |
1486 | (__v32hi) __W, | |
1487 | (__mmask32) __U); | |
1488 | } | |
1489 | ||
1490 | extern __inline __m512i | |
1491 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1492 | _mm512_maskz_unpacklo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1493 | { | |
1494 | return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, | |
1495 | (__v32hi) __B, | |
1496 | (__v32hi) | |
f0515366 | 1497 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1498 | (__mmask32) __U); |
1499 | } | |
1500 | ||
eee5d6f5 AI |
1501 | extern __inline __mmask64 |
1502 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1503 | _mm512_cmpeq_epu8_mask (__m512i __A, __m512i __B) | |
1504 | { | |
1505 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1506 | (__v64qi) __B, 0, | |
1507 | (__mmask64) -1); | |
1508 | } | |
1509 | ||
936c0fe4 AI |
1510 | extern __inline __mmask64 |
1511 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1512 | _mm512_cmpeq_epi8_mask (__m512i __A, __m512i __B) | |
1513 | { | |
1514 | return (__mmask64) __builtin_ia32_pcmpeqb512_mask ((__v64qi) __A, | |
1515 | (__v64qi) __B, | |
1516 | (__mmask64) -1); | |
1517 | } | |
1518 | ||
eee5d6f5 AI |
1519 | extern __inline __mmask64 |
1520 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1521 | _mm512_mask_cmpeq_epu8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1522 | { | |
1523 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1524 | (__v64qi) __B, 0, | |
1525 | __U); | |
1526 | } | |
1527 | ||
936c0fe4 AI |
1528 | extern __inline __mmask64 |
1529 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1530 | _mm512_mask_cmpeq_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1531 | { | |
1532 | return (__mmask64) __builtin_ia32_pcmpeqb512_mask ((__v64qi) __A, | |
1533 | (__v64qi) __B, | |
1534 | __U); | |
1535 | } | |
1536 | ||
eee5d6f5 AI |
1537 | extern __inline __mmask32 |
1538 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1539 | _mm512_cmpeq_epu16_mask (__m512i __A, __m512i __B) | |
1540 | { | |
1541 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1542 | (__v32hi) __B, 0, | |
1543 | (__mmask32) -1); | |
1544 | } | |
1545 | ||
936c0fe4 AI |
1546 | extern __inline __mmask32 |
1547 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1548 | _mm512_cmpeq_epi16_mask (__m512i __A, __m512i __B) | |
1549 | { | |
1550 | return (__mmask32) __builtin_ia32_pcmpeqw512_mask ((__v32hi) __A, | |
1551 | (__v32hi) __B, | |
1552 | (__mmask32) -1); | |
1553 | } | |
1554 | ||
eee5d6f5 AI |
1555 | extern __inline __mmask32 |
1556 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1557 | _mm512_mask_cmpeq_epu16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1558 | { | |
1559 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1560 | (__v32hi) __B, 0, | |
1561 | __U); | |
1562 | } | |
1563 | ||
936c0fe4 AI |
1564 | extern __inline __mmask32 |
1565 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1566 | _mm512_mask_cmpeq_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1567 | { | |
1568 | return (__mmask32) __builtin_ia32_pcmpeqw512_mask ((__v32hi) __A, | |
1569 | (__v32hi) __B, | |
1570 | __U); | |
1571 | } | |
1572 | ||
eee5d6f5 AI |
1573 | extern __inline __mmask64 |
1574 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1575 | _mm512_cmpgt_epu8_mask (__m512i __A, __m512i __B) | |
1576 | { | |
1577 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1578 | (__v64qi) __B, 6, | |
1579 | (__mmask64) -1); | |
1580 | } | |
1581 | ||
936c0fe4 AI |
1582 | extern __inline __mmask64 |
1583 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1584 | _mm512_cmpgt_epi8_mask (__m512i __A, __m512i __B) | |
1585 | { | |
1586 | return (__mmask64) __builtin_ia32_pcmpgtb512_mask ((__v64qi) __A, | |
1587 | (__v64qi) __B, | |
1588 | (__mmask64) -1); | |
1589 | } | |
1590 | ||
eee5d6f5 AI |
1591 | extern __inline __mmask64 |
1592 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1593 | _mm512_mask_cmpgt_epu8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1594 | { | |
1595 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1596 | (__v64qi) __B, 6, | |
1597 | __U); | |
1598 | } | |
1599 | ||
936c0fe4 AI |
1600 | extern __inline __mmask64 |
1601 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1602 | _mm512_mask_cmpgt_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1603 | { | |
1604 | return (__mmask64) __builtin_ia32_pcmpgtb512_mask ((__v64qi) __A, | |
1605 | (__v64qi) __B, | |
1606 | __U); | |
1607 | } | |
1608 | ||
eee5d6f5 AI |
1609 | extern __inline __mmask32 |
1610 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1611 | _mm512_cmpgt_epu16_mask (__m512i __A, __m512i __B) | |
1612 | { | |
1613 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1614 | (__v32hi) __B, 6, | |
1615 | (__mmask32) -1); | |
1616 | } | |
1617 | ||
936c0fe4 AI |
1618 | extern __inline __mmask32 |
1619 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1620 | _mm512_cmpgt_epi16_mask (__m512i __A, __m512i __B) | |
1621 | { | |
1622 | return (__mmask32) __builtin_ia32_pcmpgtw512_mask ((__v32hi) __A, | |
1623 | (__v32hi) __B, | |
1624 | (__mmask32) -1); | |
1625 | } | |
1626 | ||
eee5d6f5 AI |
1627 | extern __inline __mmask32 |
1628 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1629 | _mm512_mask_cmpgt_epu16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1630 | { | |
1631 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1632 | (__v32hi) __B, 6, | |
1633 | __U); | |
1634 | } | |
1635 | ||
936c0fe4 AI |
1636 | extern __inline __mmask32 |
1637 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1638 | _mm512_mask_cmpgt_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1639 | { | |
1640 | return (__mmask32) __builtin_ia32_pcmpgtw512_mask ((__v32hi) __A, | |
1641 | (__v32hi) __B, | |
1642 | __U); | |
1643 | } | |
1644 | ||
1645 | extern __inline __mmask64 | |
1646 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1647 | _mm512_movepi8_mask (__m512i __A) | |
1648 | { | |
1649 | return (__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A); | |
1650 | } | |
1651 | ||
1652 | extern __inline __mmask32 | |
1653 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1654 | _mm512_movepi16_mask (__m512i __A) | |
1655 | { | |
1656 | return (__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A); | |
1657 | } | |
1658 | ||
1659 | extern __inline __m512i | |
1660 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1661 | _mm512_movm_epi8 (__mmask64 __A) | |
1662 | { | |
1663 | return (__m512i) __builtin_ia32_cvtmask2b512 (__A); | |
1664 | } | |
1665 | ||
1666 | extern __inline __m512i | |
1667 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1668 | _mm512_movm_epi16 (__mmask32 __A) | |
1669 | { | |
1670 | return (__m512i) __builtin_ia32_cvtmask2w512 (__A); | |
1671 | } | |
1672 | ||
1673 | extern __inline __mmask64 | |
1674 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1675 | _mm512_test_epi8_mask (__m512i __A, __m512i __B) | |
1676 | { | |
1677 | return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A, | |
1678 | (__v64qi) __B, | |
1679 | (__mmask64) -1); | |
1680 | } | |
1681 | ||
1682 | extern __inline __mmask64 | |
1683 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1684 | _mm512_mask_test_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1685 | { | |
1686 | return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A, | |
1687 | (__v64qi) __B, __U); | |
1688 | } | |
1689 | ||
1690 | extern __inline __mmask32 | |
1691 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1692 | _mm512_test_epi16_mask (__m512i __A, __m512i __B) | |
1693 | { | |
1694 | return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A, | |
1695 | (__v32hi) __B, | |
1696 | (__mmask32) -1); | |
1697 | } | |
1698 | ||
1699 | extern __inline __mmask32 | |
1700 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1701 | _mm512_mask_test_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1702 | { | |
1703 | return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A, | |
1704 | (__v32hi) __B, __U); | |
1705 | } | |
1706 | ||
1707 | extern __inline __mmask64 | |
1708 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1709 | _mm512_testn_epi8_mask (__m512i __A, __m512i __B) | |
1710 | { | |
1711 | return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A, | |
1712 | (__v64qi) __B, | |
1713 | (__mmask64) -1); | |
1714 | } | |
1715 | ||
1716 | extern __inline __mmask64 | |
1717 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1718 | _mm512_mask_testn_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1719 | { | |
1720 | return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A, | |
1721 | (__v64qi) __B, __U); | |
1722 | } | |
1723 | ||
1724 | extern __inline __mmask32 | |
1725 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1726 | _mm512_testn_epi16_mask (__m512i __A, __m512i __B) | |
1727 | { | |
1728 | return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A, | |
1729 | (__v32hi) __B, | |
1730 | (__mmask32) -1); | |
1731 | } | |
1732 | ||
1733 | extern __inline __mmask32 | |
1734 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1735 | _mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1736 | { | |
1737 | return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A, | |
1738 | (__v32hi) __B, __U); | |
1739 | } | |
1740 | ||
1741 | extern __inline __m512i | |
1742 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1743 | _mm512_shuffle_epi8 (__m512i __A, __m512i __B) | |
1744 | { | |
1745 | return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, | |
1746 | (__v64qi) __B, | |
1747 | (__v64qi) | |
f0515366 | 1748 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1749 | (__mmask64) -1); |
1750 | } | |
1751 | ||
1752 | extern __inline __m512i | |
1753 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1754 | _mm512_mask_shuffle_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1755 | __m512i __B) | |
1756 | { | |
1757 | return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, | |
1758 | (__v64qi) __B, | |
1759 | (__v64qi) __W, | |
1760 | (__mmask64) __U); | |
1761 | } | |
1762 | ||
1763 | extern __inline __m512i | |
1764 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1765 | _mm512_maskz_shuffle_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1766 | { | |
1767 | return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, | |
1768 | (__v64qi) __B, | |
1769 | (__v64qi) | |
f0515366 | 1770 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1771 | (__mmask64) __U); |
1772 | } | |
1773 | ||
1774 | extern __inline __m512i | |
1775 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1776 | _mm512_min_epu16 (__m512i __A, __m512i __B) | |
1777 | { | |
1778 | return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, | |
1779 | (__v32hi) __B, | |
1780 | (__v32hi) | |
f0515366 | 1781 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1782 | (__mmask32) -1); |
1783 | } | |
1784 | ||
1785 | extern __inline __m512i | |
1786 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1787 | _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B) | |
1788 | { | |
1789 | return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, | |
1790 | (__v32hi) __B, | |
1791 | (__v32hi) | |
f0515366 | 1792 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1793 | (__mmask32) __M); |
1794 | } | |
1795 | ||
1796 | extern __inline __m512i | |
1797 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1798 | _mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, | |
1799 | __m512i __B) | |
1800 | { | |
1801 | return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, | |
1802 | (__v32hi) __B, | |
1803 | (__v32hi) __W, | |
1804 | (__mmask32) __M); | |
1805 | } | |
1806 | ||
1807 | extern __inline __m512i | |
1808 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1809 | _mm512_min_epi16 (__m512i __A, __m512i __B) | |
1810 | { | |
1811 | return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, | |
1812 | (__v32hi) __B, | |
1813 | (__v32hi) | |
f0515366 | 1814 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1815 | (__mmask32) -1); |
1816 | } | |
1817 | ||
1818 | extern __inline __m512i | |
1819 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1820 | _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B) | |
1821 | { | |
1822 | return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, | |
1823 | (__v32hi) __B, | |
1824 | (__v32hi) | |
f0515366 | 1825 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1826 | (__mmask32) __M); |
1827 | } | |
1828 | ||
1829 | extern __inline __m512i | |
1830 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1831 | _mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, | |
1832 | __m512i __B) | |
1833 | { | |
1834 | return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, | |
1835 | (__v32hi) __B, | |
1836 | (__v32hi) __W, | |
1837 | (__mmask32) __M); | |
1838 | } | |
1839 | ||
1840 | extern __inline __m512i | |
1841 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1842 | _mm512_max_epu8 (__m512i __A, __m512i __B) | |
1843 | { | |
1844 | return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, | |
1845 | (__v64qi) __B, | |
1846 | (__v64qi) | |
f0515366 | 1847 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1848 | (__mmask64) -1); |
1849 | } | |
1850 | ||
1851 | extern __inline __m512i | |
1852 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1853 | _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B) | |
1854 | { | |
1855 | return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, | |
1856 | (__v64qi) __B, | |
1857 | (__v64qi) | |
f0515366 | 1858 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1859 | (__mmask64) __M); |
1860 | } | |
1861 | ||
1862 | extern __inline __m512i | |
1863 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1864 | _mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, | |
1865 | __m512i __B) | |
1866 | { | |
1867 | return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, | |
1868 | (__v64qi) __B, | |
1869 | (__v64qi) __W, | |
1870 | (__mmask64) __M); | |
1871 | } | |
1872 | ||
1873 | extern __inline __m512i | |
1874 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1875 | _mm512_max_epi8 (__m512i __A, __m512i __B) | |
1876 | { | |
1877 | return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, | |
1878 | (__v64qi) __B, | |
1879 | (__v64qi) | |
f0515366 | 1880 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1881 | (__mmask64) -1); |
1882 | } | |
1883 | ||
1884 | extern __inline __m512i | |
1885 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1886 | _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B) | |
1887 | { | |
1888 | return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, | |
1889 | (__v64qi) __B, | |
1890 | (__v64qi) | |
f0515366 | 1891 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1892 | (__mmask64) __M); |
1893 | } | |
1894 | ||
1895 | extern __inline __m512i | |
1896 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1897 | _mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, | |
1898 | __m512i __B) | |
1899 | { | |
1900 | return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, | |
1901 | (__v64qi) __B, | |
1902 | (__v64qi) __W, | |
1903 | (__mmask64) __M); | |
1904 | } | |
1905 | ||
1906 | extern __inline __m512i | |
1907 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1908 | _mm512_min_epu8 (__m512i __A, __m512i __B) | |
1909 | { | |
1910 | return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, | |
1911 | (__v64qi) __B, | |
1912 | (__v64qi) | |
f0515366 | 1913 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1914 | (__mmask64) -1); |
1915 | } | |
1916 | ||
1917 | extern __inline __m512i | |
1918 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1919 | _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B) | |
1920 | { | |
1921 | return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, | |
1922 | (__v64qi) __B, | |
1923 | (__v64qi) | |
f0515366 | 1924 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1925 | (__mmask64) __M); |
1926 | } | |
1927 | ||
1928 | extern __inline __m512i | |
1929 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1930 | _mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, | |
1931 | __m512i __B) | |
1932 | { | |
1933 | return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, | |
1934 | (__v64qi) __B, | |
1935 | (__v64qi) __W, | |
1936 | (__mmask64) __M); | |
1937 | } | |
1938 | ||
1939 | extern __inline __m512i | |
1940 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1941 | _mm512_min_epi8 (__m512i __A, __m512i __B) | |
1942 | { | |
1943 | return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, | |
1944 | (__v64qi) __B, | |
1945 | (__v64qi) | |
f0515366 | 1946 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1947 | (__mmask64) -1); |
1948 | } | |
1949 | ||
1950 | extern __inline __m512i | |
1951 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1952 | _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B) | |
1953 | { | |
1954 | return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, | |
1955 | (__v64qi) __B, | |
1956 | (__v64qi) | |
f0515366 | 1957 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1958 | (__mmask64) __M); |
1959 | } | |
1960 | ||
1961 | extern __inline __m512i | |
1962 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1963 | _mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, | |
1964 | __m512i __B) | |
1965 | { | |
1966 | return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, | |
1967 | (__v64qi) __B, | |
1968 | (__v64qi) __W, | |
1969 | (__mmask64) __M); | |
1970 | } | |
1971 | ||
1972 | extern __inline __m512i | |
1973 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1974 | _mm512_max_epi16 (__m512i __A, __m512i __B) | |
1975 | { | |
1976 | return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, | |
1977 | (__v32hi) __B, | |
1978 | (__v32hi) | |
f0515366 | 1979 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1980 | (__mmask32) -1); |
1981 | } | |
1982 | ||
1983 | extern __inline __m512i | |
1984 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1985 | _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B) | |
1986 | { | |
1987 | return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, | |
1988 | (__v32hi) __B, | |
1989 | (__v32hi) | |
f0515366 | 1990 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1991 | (__mmask32) __M); |
1992 | } | |
1993 | ||
1994 | extern __inline __m512i | |
1995 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1996 | _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A, | |
1997 | __m512i __B) | |
1998 | { | |
1999 | return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, | |
2000 | (__v32hi) __B, | |
2001 | (__v32hi) __W, | |
2002 | (__mmask32) __M); | |
2003 | } | |
2004 | ||
2005 | extern __inline __m512i | |
2006 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2007 | _mm512_max_epu16 (__m512i __A, __m512i __B) | |
2008 | { | |
2009 | return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, | |
2010 | (__v32hi) __B, | |
2011 | (__v32hi) | |
f0515366 | 2012 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2013 | (__mmask32) -1); |
2014 | } | |
2015 | ||
2016 | extern __inline __m512i | |
2017 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2018 | _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B) | |
2019 | { | |
2020 | return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, | |
2021 | (__v32hi) __B, | |
2022 | (__v32hi) | |
f0515366 | 2023 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2024 | (__mmask32) __M); |
2025 | } | |
2026 | ||
2027 | extern __inline __m512i | |
2028 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2029 | _mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, | |
2030 | __m512i __B) | |
2031 | { | |
2032 | return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, | |
2033 | (__v32hi) __B, | |
2034 | (__v32hi) __W, | |
2035 | (__mmask32) __M); | |
2036 | } | |
2037 | ||
2038 | extern __inline __m512i | |
2039 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2040 | _mm512_sra_epi16 (__m512i __A, __m128i __B) | |
2041 | { | |
2042 | return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, | |
2043 | (__v8hi) __B, | |
2044 | (__v32hi) | |
f0515366 | 2045 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2046 | (__mmask32) -1); |
2047 | } | |
2048 | ||
2049 | extern __inline __m512i | |
2050 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2051 | _mm512_mask_sra_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2052 | __m128i __B) | |
2053 | { | |
2054 | return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, | |
2055 | (__v8hi) __B, | |
2056 | (__v32hi) __W, | |
2057 | (__mmask32) __U); | |
2058 | } | |
2059 | ||
2060 | extern __inline __m512i | |
2061 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2062 | _mm512_maskz_sra_epi16 (__mmask32 __U, __m512i __A, __m128i __B) | |
2063 | { | |
2064 | return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, | |
2065 | (__v8hi) __B, | |
2066 | (__v32hi) | |
f0515366 | 2067 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2068 | (__mmask32) __U); |
2069 | } | |
2070 | ||
2071 | extern __inline __m512i | |
2072 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2073 | _mm512_srav_epi16 (__m512i __A, __m512i __B) | |
2074 | { | |
2075 | return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, | |
2076 | (__v32hi) __B, | |
2077 | (__v32hi) | |
f0515366 | 2078 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2079 | (__mmask32) -1); |
2080 | } | |
2081 | ||
2082 | extern __inline __m512i | |
2083 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2084 | _mm512_mask_srav_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2085 | __m512i __B) | |
2086 | { | |
2087 | return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, | |
2088 | (__v32hi) __B, | |
2089 | (__v32hi) __W, | |
2090 | (__mmask32) __U); | |
2091 | } | |
2092 | ||
2093 | extern __inline __m512i | |
2094 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2095 | _mm512_maskz_srav_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
2096 | { | |
2097 | return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, | |
2098 | (__v32hi) __B, | |
2099 | (__v32hi) | |
f0515366 | 2100 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2101 | (__mmask32) __U); |
2102 | } | |
2103 | ||
2104 | extern __inline __m512i | |
2105 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2106 | _mm512_srlv_epi16 (__m512i __A, __m512i __B) | |
2107 | { | |
2108 | return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, | |
2109 | (__v32hi) __B, | |
2110 | (__v32hi) | |
f0515366 | 2111 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2112 | (__mmask32) -1); |
2113 | } | |
2114 | ||
2115 | extern __inline __m512i | |
2116 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2117 | _mm512_mask_srlv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2118 | __m512i __B) | |
2119 | { | |
2120 | return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, | |
2121 | (__v32hi) __B, | |
2122 | (__v32hi) __W, | |
2123 | (__mmask32) __U); | |
2124 | } | |
2125 | ||
2126 | extern __inline __m512i | |
2127 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2128 | _mm512_maskz_srlv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
2129 | { | |
2130 | return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, | |
2131 | (__v32hi) __B, | |
2132 | (__v32hi) | |
f0515366 | 2133 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2134 | (__mmask32) __U); |
2135 | } | |
2136 | ||
2137 | extern __inline __m512i | |
2138 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2139 | _mm512_sllv_epi16 (__m512i __A, __m512i __B) | |
2140 | { | |
2141 | return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, | |
2142 | (__v32hi) __B, | |
2143 | (__v32hi) | |
f0515366 | 2144 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2145 | (__mmask32) -1); |
2146 | } | |
2147 | ||
2148 | extern __inline __m512i | |
2149 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2150 | _mm512_mask_sllv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2151 | __m512i __B) | |
2152 | { | |
2153 | return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, | |
2154 | (__v32hi) __B, | |
2155 | (__v32hi) __W, | |
2156 | (__mmask32) __U); | |
2157 | } | |
2158 | ||
2159 | extern __inline __m512i | |
2160 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2161 | _mm512_maskz_sllv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
2162 | { | |
2163 | return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, | |
2164 | (__v32hi) __B, | |
2165 | (__v32hi) | |
f0515366 | 2166 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2167 | (__mmask32) __U); |
2168 | } | |
2169 | ||
2170 | extern __inline __m512i | |
2171 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2172 | _mm512_mask_packs_epi16 (__m512i __W, __mmask64 __M, __m512i __A, | |
2173 | __m512i __B) | |
2174 | { | |
2175 | return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, | |
2176 | (__v32hi) __B, | |
2177 | (__v64qi) __W, | |
2178 | (__mmask64) __M); | |
2179 | } | |
2180 | ||
2181 | extern __inline __m512i | |
2182 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2183 | _mm512_maskz_packs_epi16 (__mmask64 __M, __m512i __A, __m512i __B) | |
2184 | { | |
2185 | return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, | |
2186 | (__v32hi) __B, | |
2187 | (__v64qi) | |
f0515366 | 2188 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2189 | __M); |
2190 | } | |
2191 | ||
2192 | extern __inline __m512i | |
2193 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2194 | _mm512_packus_epi16 (__m512i __A, __m512i __B) | |
2195 | { | |
2196 | return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, | |
2197 | (__v32hi) __B, | |
2198 | (__v64qi) | |
f0515366 | 2199 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2200 | (__mmask64) -1); |
2201 | } | |
2202 | ||
2203 | extern __inline __m512i | |
2204 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2205 | _mm512_mask_packus_epi16 (__m512i __W, __mmask64 __M, __m512i __A, | |
2206 | __m512i __B) | |
2207 | { | |
2208 | return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, | |
2209 | (__v32hi) __B, | |
2210 | (__v64qi) __W, | |
2211 | (__mmask64) __M); | |
2212 | } | |
2213 | ||
2214 | extern __inline __m512i | |
2215 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2216 | _mm512_maskz_packus_epi16 (__mmask64 __M, __m512i __A, __m512i __B) | |
2217 | { | |
2218 | return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, | |
2219 | (__v32hi) __B, | |
2220 | (__v64qi) | |
f0515366 | 2221 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2222 | (__mmask64) __M); |
2223 | } | |
2224 | ||
2225 | extern __inline __m512i | |
2226 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2227 | _mm512_abs_epi8 (__m512i __A) | |
2228 | { | |
2229 | return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, | |
2230 | (__v64qi) | |
f0515366 | 2231 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2232 | (__mmask64) -1); |
2233 | } | |
2234 | ||
2235 | extern __inline __m512i | |
2236 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2237 | _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A) | |
2238 | { | |
2239 | return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, | |
2240 | (__v64qi) __W, | |
2241 | (__mmask64) __U); | |
2242 | } | |
2243 | ||
2244 | extern __inline __m512i | |
2245 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2246 | _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A) | |
2247 | { | |
2248 | return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, | |
2249 | (__v64qi) | |
f0515366 | 2250 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2251 | (__mmask64) __U); |
2252 | } | |
2253 | ||
2254 | extern __inline __m512i | |
2255 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2256 | _mm512_abs_epi16 (__m512i __A) | |
2257 | { | |
2258 | return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, | |
2259 | (__v32hi) | |
f0515366 | 2260 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2261 | (__mmask32) -1); |
2262 | } | |
2263 | ||
2264 | extern __inline __m512i | |
2265 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2266 | _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A) | |
2267 | { | |
2268 | return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, | |
2269 | (__v32hi) __W, | |
2270 | (__mmask32) __U); | |
2271 | } | |
2272 | ||
2273 | extern __inline __m512i | |
2274 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2275 | _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A) | |
2276 | { | |
2277 | return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, | |
2278 | (__v32hi) | |
f0515366 | 2279 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2280 | (__mmask32) __U); |
2281 | } | |
2282 | ||
eee5d6f5 AI |
2283 | extern __inline __mmask64 |
2284 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2285 | _mm512_mask_cmpneq_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2286 | { | |
2287 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2288 | (__v64qi) __Y, 4, | |
2289 | (__mmask64) __M); | |
2290 | } | |
2291 | ||
2292 | extern __inline __mmask64 | |
2293 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2294 | _mm512_mask_cmplt_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2295 | { | |
2296 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2297 | (__v64qi) __Y, 1, | |
2298 | (__mmask64) __M); | |
2299 | } | |
2300 | ||
2301 | extern __inline __mmask64 | |
2302 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2303 | _mm512_mask_cmpge_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2304 | { | |
2305 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2306 | (__v64qi) __Y, 5, | |
2307 | (__mmask64) __M); | |
2308 | } | |
2309 | ||
2310 | extern __inline __mmask64 | |
2311 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2312 | _mm512_mask_cmple_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2313 | { | |
2314 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2315 | (__v64qi) __Y, 2, | |
2316 | (__mmask64) __M); | |
2317 | } | |
2318 | ||
2319 | extern __inline __mmask32 | |
2320 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2321 | _mm512_mask_cmpneq_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2322 | { | |
2323 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2324 | (__v32hi) __Y, 4, | |
2325 | (__mmask32) __M); | |
2326 | } | |
2327 | ||
2328 | extern __inline __mmask32 | |
2329 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2330 | _mm512_mask_cmplt_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2331 | { | |
2332 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2333 | (__v32hi) __Y, 1, | |
2334 | (__mmask32) __M); | |
2335 | } | |
2336 | ||
2337 | extern __inline __mmask32 | |
2338 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2339 | _mm512_mask_cmpge_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2340 | { | |
2341 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2342 | (__v32hi) __Y, 5, | |
2343 | (__mmask32) __M); | |
2344 | } | |
2345 | ||
2346 | extern __inline __mmask32 | |
2347 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2348 | _mm512_mask_cmple_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2349 | { | |
2350 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2351 | (__v32hi) __Y, 2, | |
2352 | (__mmask32) __M); | |
2353 | } | |
2354 | ||
2355 | extern __inline __mmask64 | |
2356 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2357 | _mm512_mask_cmpneq_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2358 | { | |
2359 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2360 | (__v64qi) __Y, 4, | |
2361 | (__mmask64) __M); | |
2362 | } | |
2363 | ||
2364 | extern __inline __mmask64 | |
2365 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2366 | _mm512_mask_cmplt_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2367 | { | |
2368 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2369 | (__v64qi) __Y, 1, | |
2370 | (__mmask64) __M); | |
2371 | } | |
2372 | ||
2373 | extern __inline __mmask64 | |
2374 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2375 | _mm512_mask_cmpge_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2376 | { | |
2377 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2378 | (__v64qi) __Y, 5, | |
2379 | (__mmask64) __M); | |
2380 | } | |
2381 | ||
2382 | extern __inline __mmask64 | |
2383 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2384 | _mm512_mask_cmple_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2385 | { | |
2386 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2387 | (__v64qi) __Y, 2, | |
2388 | (__mmask64) __M); | |
2389 | } | |
2390 | ||
2391 | extern __inline __mmask32 | |
2392 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2393 | _mm512_mask_cmpneq_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2394 | { | |
2395 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2396 | (__v32hi) __Y, 4, | |
2397 | (__mmask32) __M); | |
2398 | } | |
2399 | ||
2400 | extern __inline __mmask32 | |
2401 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2402 | _mm512_mask_cmplt_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2403 | { | |
2404 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2405 | (__v32hi) __Y, 1, | |
2406 | (__mmask32) __M); | |
2407 | } | |
2408 | ||
2409 | extern __inline __mmask32 | |
2410 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2411 | _mm512_mask_cmpge_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2412 | { | |
2413 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2414 | (__v32hi) __Y, 5, | |
2415 | (__mmask32) __M); | |
2416 | } | |
2417 | ||
2418 | extern __inline __mmask32 | |
2419 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2420 | _mm512_mask_cmple_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2421 | { | |
2422 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2423 | (__v32hi) __Y, 2, | |
2424 | (__mmask32) __M); | |
2425 | } | |
2426 | ||
2427 | extern __inline __mmask64 | |
2428 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2429 | _mm512_cmpneq_epu8_mask (__m512i __X, __m512i __Y) | |
2430 | { | |
2431 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2432 | (__v64qi) __Y, 4, | |
281f7120 | 2433 | (__mmask64) -1); |
eee5d6f5 AI |
2434 | } |
2435 | ||
2436 | extern __inline __mmask64 | |
2437 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2438 | _mm512_cmplt_epu8_mask (__m512i __X, __m512i __Y) | |
2439 | { | |
2440 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2441 | (__v64qi) __Y, 1, | |
281f7120 | 2442 | (__mmask64) -1); |
eee5d6f5 AI |
2443 | } |
2444 | ||
2445 | extern __inline __mmask64 | |
2446 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2447 | _mm512_cmpge_epu8_mask (__m512i __X, __m512i __Y) | |
2448 | { | |
2449 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2450 | (__v64qi) __Y, 5, | |
281f7120 | 2451 | (__mmask64) -1); |
eee5d6f5 AI |
2452 | } |
2453 | ||
2454 | extern __inline __mmask64 | |
2455 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2456 | _mm512_cmple_epu8_mask (__m512i __X, __m512i __Y) | |
2457 | { | |
2458 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2459 | (__v64qi) __Y, 2, | |
281f7120 | 2460 | (__mmask64) -1); |
eee5d6f5 AI |
2461 | } |
2462 | ||
2463 | extern __inline __mmask32 | |
2464 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2465 | _mm512_cmpneq_epu16_mask (__m512i __X, __m512i __Y) | |
2466 | { | |
2467 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2468 | (__v32hi) __Y, 4, | |
281f7120 | 2469 | (__mmask32) -1); |
eee5d6f5 AI |
2470 | } |
2471 | ||
2472 | extern __inline __mmask32 | |
2473 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2474 | _mm512_cmplt_epu16_mask (__m512i __X, __m512i __Y) | |
2475 | { | |
2476 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2477 | (__v32hi) __Y, 1, | |
281f7120 | 2478 | (__mmask32) -1); |
eee5d6f5 AI |
2479 | } |
2480 | ||
2481 | extern __inline __mmask32 | |
2482 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2483 | _mm512_cmpge_epu16_mask (__m512i __X, __m512i __Y) | |
2484 | { | |
2485 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2486 | (__v32hi) __Y, 5, | |
281f7120 | 2487 | (__mmask32) -1); |
eee5d6f5 AI |
2488 | } |
2489 | ||
2490 | extern __inline __mmask32 | |
2491 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2492 | _mm512_cmple_epu16_mask (__m512i __X, __m512i __Y) | |
2493 | { | |
2494 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2495 | (__v32hi) __Y, 2, | |
281f7120 | 2496 | (__mmask32) -1); |
eee5d6f5 AI |
2497 | } |
2498 | ||
2499 | extern __inline __mmask64 | |
2500 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2501 | _mm512_cmpneq_epi8_mask (__m512i __X, __m512i __Y) | |
2502 | { | |
2503 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2504 | (__v64qi) __Y, 4, | |
281f7120 | 2505 | (__mmask64) -1); |
eee5d6f5 AI |
2506 | } |
2507 | ||
2508 | extern __inline __mmask64 | |
2509 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2510 | _mm512_cmplt_epi8_mask (__m512i __X, __m512i __Y) | |
2511 | { | |
2512 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2513 | (__v64qi) __Y, 1, | |
281f7120 | 2514 | (__mmask64) -1); |
eee5d6f5 AI |
2515 | } |
2516 | ||
2517 | extern __inline __mmask64 | |
2518 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2519 | _mm512_cmpge_epi8_mask (__m512i __X, __m512i __Y) | |
2520 | { | |
2521 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2522 | (__v64qi) __Y, 5, | |
281f7120 | 2523 | (__mmask64) -1); |
eee5d6f5 AI |
2524 | } |
2525 | ||
2526 | extern __inline __mmask64 | |
2527 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2528 | _mm512_cmple_epi8_mask (__m512i __X, __m512i __Y) | |
2529 | { | |
2530 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2531 | (__v64qi) __Y, 2, | |
281f7120 | 2532 | (__mmask64) -1); |
eee5d6f5 AI |
2533 | } |
2534 | ||
2535 | extern __inline __mmask32 | |
2536 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2537 | _mm512_cmpneq_epi16_mask (__m512i __X, __m512i __Y) | |
2538 | { | |
2539 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2540 | (__v32hi) __Y, 4, | |
281f7120 | 2541 | (__mmask32) -1); |
eee5d6f5 AI |
2542 | } |
2543 | ||
2544 | extern __inline __mmask32 | |
2545 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2546 | _mm512_cmplt_epi16_mask (__m512i __X, __m512i __Y) | |
2547 | { | |
2548 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2549 | (__v32hi) __Y, 1, | |
281f7120 | 2550 | (__mmask32) -1); |
eee5d6f5 AI |
2551 | } |
2552 | ||
2553 | extern __inline __mmask32 | |
2554 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2555 | _mm512_cmpge_epi16_mask (__m512i __X, __m512i __Y) | |
2556 | { | |
2557 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2558 | (__v32hi) __Y, 5, | |
281f7120 | 2559 | (__mmask32) -1); |
eee5d6f5 AI |
2560 | } |
2561 | ||
2562 | extern __inline __mmask32 | |
2563 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2564 | _mm512_cmple_epi16_mask (__m512i __X, __m512i __Y) | |
2565 | { | |
2566 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2567 | (__v32hi) __Y, 2, | |
281f7120 | 2568 | (__mmask32) -1); |
eee5d6f5 AI |
2569 | } |
2570 | ||
936c0fe4 AI |
2571 | #ifdef __OPTIMIZE__ |
2572 | extern __inline __m512i | |
2573 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2574 | _mm512_alignr_epi8 (__m512i __A, __m512i __B, const int __N) | |
2575 | { | |
2576 | return (__m512i) __builtin_ia32_palignr512 ((__v8di) __A, | |
2577 | (__v8di) __B, __N * 8); | |
2578 | } | |
2579 | ||
2580 | extern __inline __m512i | |
2581 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2582 | _mm512_mask_alignr_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
2583 | __m512i __B, const int __N) | |
2584 | { | |
2585 | return (__m512i) __builtin_ia32_palignr512_mask ((__v8di) __A, | |
2586 | (__v8di) __B, | |
2587 | __N * 8, | |
2588 | (__v8di) __W, | |
2589 | (__mmask64) __U); | |
2590 | } | |
2591 | ||
2592 | extern __inline __m512i | |
2593 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2594 | _mm512_maskz_alignr_epi8 (__mmask64 __U, __m512i __A, __m512i __B, | |
2595 | const int __N) | |
2596 | { | |
2597 | return (__m512i) __builtin_ia32_palignr512_mask ((__v8di) __A, | |
2598 | (__v8di) __B, | |
2599 | __N * 8, | |
2600 | (__v8di) | |
2601 | _mm512_setzero_si512 (), | |
2602 | (__mmask64) __U); | |
2603 | } | |
2604 | ||
2605 | extern __inline __m512i | |
2606 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2607 | _mm512_dbsad_epu8 (__m512i __A, __m512i __B, const int __imm) | |
2608 | { | |
2609 | return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, | |
2610 | (__v64qi) __B, | |
2611 | __imm, | |
2612 | (__v32hi) | |
f0515366 | 2613 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2614 | (__mmask32) -1); |
2615 | } | |
2616 | ||
2617 | extern __inline __m512i | |
2618 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2619 | _mm512_mask_dbsad_epu8 (__m512i __W, __mmask32 __U, __m512i __A, | |
2620 | __m512i __B, const int __imm) | |
2621 | { | |
2622 | return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, | |
2623 | (__v64qi) __B, | |
2624 | __imm, | |
2625 | (__v32hi) __W, | |
2626 | (__mmask32) __U); | |
2627 | } | |
2628 | ||
2629 | extern __inline __m512i | |
2630 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2631 | _mm512_maskz_dbsad_epu8 (__mmask32 __U, __m512i __A, __m512i __B, | |
2632 | const int __imm) | |
2633 | { | |
2634 | return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, | |
2635 | (__v64qi) __B, | |
2636 | __imm, | |
2637 | (__v32hi) | |
f0515366 | 2638 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2639 | (__mmask32) __U); |
2640 | } | |
2641 | ||
2642 | extern __inline __m512i | |
2643 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2644 | _mm512_srli_epi16 (__m512i __A, const int __imm) | |
2645 | { | |
2646 | return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, | |
2647 | (__v32hi) | |
f0515366 | 2648 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2649 | (__mmask32) -1); |
2650 | } | |
2651 | ||
2652 | extern __inline __m512i | |
2653 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2654 | _mm512_mask_srli_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2655 | const int __imm) | |
2656 | { | |
2657 | return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, | |
2658 | (__v32hi) __W, | |
2659 | (__mmask32) __U); | |
2660 | } | |
2661 | ||
2662 | extern __inline __m512i | |
2663 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2664 | _mm512_maskz_srli_epi16 (__mmask32 __U, __m512i __A, const int __imm) | |
2665 | { | |
2666 | return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, | |
2667 | (__v32hi) | |
f0515366 | 2668 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2669 | (__mmask32) __U); |
2670 | } | |
2671 | ||
2672 | extern __inline __m512i | |
2673 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2674 | _mm512_slli_epi16 (__m512i __A, const int __B) | |
2675 | { | |
2676 | return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, | |
2677 | (__v32hi) | |
f0515366 | 2678 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2679 | (__mmask32) -1); |
2680 | } | |
2681 | ||
2682 | extern __inline __m512i | |
2683 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2684 | _mm512_mask_slli_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2685 | const int __B) | |
2686 | { | |
2687 | return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, | |
2688 | (__v32hi) __W, | |
2689 | (__mmask32) __U); | |
2690 | } | |
2691 | ||
2692 | extern __inline __m512i | |
2693 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2694 | _mm512_maskz_slli_epi16 (__mmask32 __U, __m512i __A, const int __B) | |
2695 | { | |
2696 | return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, | |
2697 | (__v32hi) | |
f0515366 | 2698 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2699 | (__mmask32) __U); |
2700 | } | |
2701 | ||
2702 | extern __inline __m512i | |
2703 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2704 | _mm512_shufflehi_epi16 (__m512i __A, const int __imm) | |
2705 | { | |
2706 | return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, | |
2707 | __imm, | |
2708 | (__v32hi) | |
f0515366 | 2709 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2710 | (__mmask32) -1); |
2711 | } | |
2712 | ||
2713 | extern __inline __m512i | |
2714 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2715 | _mm512_mask_shufflehi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2716 | const int __imm) | |
2717 | { | |
2718 | return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, | |
2719 | __imm, | |
2720 | (__v32hi) __W, | |
2721 | (__mmask32) __U); | |
2722 | } | |
2723 | ||
2724 | extern __inline __m512i | |
2725 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2726 | _mm512_maskz_shufflehi_epi16 (__mmask32 __U, __m512i __A, | |
2727 | const int __imm) | |
2728 | { | |
2729 | return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, | |
2730 | __imm, | |
2731 | (__v32hi) | |
f0515366 | 2732 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2733 | (__mmask32) __U); |
2734 | } | |
2735 | ||
2736 | extern __inline __m512i | |
2737 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2738 | _mm512_shufflelo_epi16 (__m512i __A, const int __imm) | |
2739 | { | |
2740 | return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, | |
2741 | __imm, | |
2742 | (__v32hi) | |
f0515366 | 2743 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2744 | (__mmask32) -1); |
2745 | } | |
2746 | ||
2747 | extern __inline __m512i | |
2748 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2749 | _mm512_mask_shufflelo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2750 | const int __imm) | |
2751 | { | |
2752 | return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, | |
2753 | __imm, | |
2754 | (__v32hi) __W, | |
2755 | (__mmask32) __U); | |
2756 | } | |
2757 | ||
2758 | extern __inline __m512i | |
2759 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2760 | _mm512_maskz_shufflelo_epi16 (__mmask32 __U, __m512i __A, | |
2761 | const int __imm) | |
2762 | { | |
2763 | return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, | |
2764 | __imm, | |
2765 | (__v32hi) | |
f0515366 | 2766 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2767 | (__mmask32) __U); |
2768 | } | |
2769 | ||
2770 | extern __inline __m512i | |
2771 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2772 | _mm512_srai_epi16 (__m512i __A, const int __imm) | |
2773 | { | |
2774 | return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, | |
2775 | (__v32hi) | |
f0515366 | 2776 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2777 | (__mmask32) -1); |
2778 | } | |
2779 | ||
2780 | extern __inline __m512i | |
2781 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2782 | _mm512_mask_srai_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2783 | const int __imm) | |
2784 | { | |
2785 | return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, | |
2786 | (__v32hi) __W, | |
2787 | (__mmask32) __U); | |
2788 | } | |
2789 | ||
2790 | extern __inline __m512i | |
2791 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2792 | _mm512_maskz_srai_epi16 (__mmask32 __U, __m512i __A, const int __imm) | |
2793 | { | |
2794 | return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, | |
2795 | (__v32hi) | |
f0515366 | 2796 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2797 | (__mmask32) __U); |
2798 | } | |
2799 | ||
2800 | extern __inline __m512i | |
2801 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2802 | _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W) | |
2803 | { | |
2804 | return (__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) __A, | |
2805 | (__v32hi) __W, | |
2806 | (__mmask32) __U); | |
2807 | } | |
2808 | ||
2809 | extern __inline __m512i | |
2810 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2811 | _mm512_mask_blend_epi8 (__mmask64 __U, __m512i __A, __m512i __W) | |
2812 | { | |
2813 | return (__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) __A, | |
2814 | (__v64qi) __W, | |
2815 | (__mmask64) __U); | |
2816 | } | |
2817 | ||
2818 | extern __inline __mmask32 | |
2819 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2820 | _mm512_mask_cmp_epi16_mask (__mmask32 __U, __m512i __X, __m512i __Y, | |
2821 | const int __P) | |
2822 | { | |
2823 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2824 | (__v32hi) __Y, __P, | |
2825 | (__mmask32) __U); | |
2826 | } | |
2827 | ||
2828 | extern __inline __mmask32 | |
2829 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2830 | _mm512_cmp_epi16_mask (__m512i __X, __m512i __Y, const int __P) | |
2831 | { | |
2832 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2833 | (__v32hi) __Y, __P, | |
2834 | (__mmask32) -1); | |
2835 | } | |
2836 | ||
2837 | extern __inline __mmask64 | |
2838 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2839 | _mm512_mask_cmp_epi8_mask (__mmask32 __U, __m512i __X, __m512i __Y, | |
2840 | const int __P) | |
2841 | { | |
2842 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2843 | (__v64qi) __Y, __P, | |
2844 | (__mmask64) __U); | |
2845 | } | |
2846 | ||
2847 | extern __inline __mmask64 | |
2848 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2849 | _mm512_cmp_epi8_mask (__m512i __X, __m512i __Y, const int __P) | |
2850 | { | |
2851 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2852 | (__v64qi) __Y, __P, | |
2853 | (__mmask64) -1); | |
2854 | } | |
2855 | ||
2856 | extern __inline __mmask32 | |
2857 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2858 | _mm512_mask_cmp_epu16_mask (__mmask32 __U, __m512i __X, __m512i __Y, | |
2859 | const int __P) | |
2860 | { | |
2861 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2862 | (__v32hi) __Y, __P, | |
2863 | (__mmask32) __U); | |
2864 | } | |
2865 | ||
2866 | extern __inline __mmask32 | |
2867 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2868 | _mm512_cmp_epu16_mask (__m512i __X, __m512i __Y, const int __P) | |
2869 | { | |
2870 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2871 | (__v32hi) __Y, __P, | |
2872 | (__mmask32) -1); | |
2873 | } | |
2874 | ||
2875 | extern __inline __mmask64 | |
2876 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2877 | _mm512_mask_cmp_epu8_mask (__mmask32 __U, __m512i __X, __m512i __Y, | |
2878 | const int __P) | |
2879 | { | |
2880 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2881 | (__v64qi) __Y, __P, | |
2882 | (__mmask64) __U); | |
2883 | } | |
2884 | ||
2885 | extern __inline __mmask64 | |
2886 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2887 | _mm512_cmp_epu8_mask (__m512i __X, __m512i __Y, const int __P) | |
2888 | { | |
2889 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2890 | (__v64qi) __Y, __P, | |
2891 | (__mmask64) -1); | |
2892 | } | |
2893 | ||
2894 | extern __inline __m512i | |
2895 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2896 | _mm512_packs_epi32 (__m512i __A, __m512i __B) | |
2897 | { | |
2898 | return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, | |
2899 | (__v16si) __B, | |
2900 | (__v32hi) | |
f0515366 | 2901 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2902 | (__mmask32) -1); |
2903 | } | |
2904 | ||
2905 | extern __inline __m512i | |
2906 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2907 | _mm512_maskz_packs_epi32 (__mmask32 __M, __m512i __A, __m512i __B) | |
2908 | { | |
2909 | return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, | |
2910 | (__v16si) __B, | |
2911 | (__v32hi) | |
f0515366 | 2912 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2913 | __M); |
2914 | } | |
2915 | ||
2916 | extern __inline __m512i | |
2917 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2918 | _mm512_mask_packs_epi32 (__m512i __W, __mmask32 __M, __m512i __A, | |
2919 | __m512i __B) | |
2920 | { | |
2921 | return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, | |
2922 | (__v16si) __B, | |
2923 | (__v32hi) __W, | |
2924 | __M); | |
2925 | } | |
2926 | ||
2927 | extern __inline __m512i | |
2928 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2929 | _mm512_packus_epi32 (__m512i __A, __m512i __B) | |
2930 | { | |
2931 | return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, | |
2932 | (__v16si) __B, | |
2933 | (__v32hi) | |
f0515366 | 2934 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2935 | (__mmask32) -1); |
2936 | } | |
2937 | ||
2938 | extern __inline __m512i | |
2939 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2940 | _mm512_maskz_packus_epi32 (__mmask32 __M, __m512i __A, __m512i __B) | |
2941 | { | |
2942 | return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, | |
2943 | (__v16si) __B, | |
2944 | (__v32hi) | |
f0515366 | 2945 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2946 | __M); |
2947 | } | |
2948 | ||
2949 | extern __inline __m512i | |
2950 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2951 | _mm512_mask_packus_epi32 (__m512i __W, __mmask32 __M, __m512i __A, | |
2952 | __m512i __B) | |
2953 | { | |
2954 | return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, | |
2955 | (__v16si) __B, | |
2956 | (__v32hi) __W, | |
2957 | __M); | |
2958 | } | |
2959 | ||
2960 | extern __inline __m512i | |
2961 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2962 | _mm512_bslli_epi128 (__m512i __A, const int __N) | |
2963 | { | |
2964 | return (__m512i) __builtin_ia32_pslldq512 (__A, __N * 8); | |
2965 | } | |
2966 | ||
2967 | extern __inline __m512i | |
2968 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2969 | _mm512_bsrli_epi128 (__m512i __A, const int __N) | |
2970 | { | |
2971 | return (__m512i) __builtin_ia32_psrldq512 (__A, __N * 8); | |
2972 | } | |
2973 | ||
2974 | #else | |
2975 | #define _mm512_alignr_epi8(X, Y, N) \ | |
2976 | ((__m512i) __builtin_ia32_palignr512 ((__v8di)(__m512i)(X), \ | |
2977 | (__v8di)(__m512i)(Y), \ | |
2978 | (int)(N * 8))) | |
2979 | ||
2980 | #define _mm512_mask_alignr_epi8(W, U, X, Y, N) \ | |
2981 | ((__m512i) __builtin_ia32_palignr512_mask ((__v8di)(__m512i)(X), \ | |
2982 | (__v8di)(__m512i)(Y), (int)(N * 8), \ | |
2983 | (__v8di)(__m512i)(W), (__mmask64)(U))) | |
2984 | ||
2985 | #define _mm512_maskz_alignr_epi8(U, X, Y, N) \ | |
2986 | ((__m512i) __builtin_ia32_palignr512_mask ((__v8di)(__m512i)(X), \ | |
f0515366 JJ |
2987 | (__v8di)(__m512i)(Y), (int)(N * 8), \ |
2988 | (__v8di)(__m512i) \ | |
2989 | _mm512_setzero_si512 (), \ | |
2990 | (__mmask64)(U))) | |
936c0fe4 AI |
2991 | |
2992 | #define _mm512_dbsad_epu8(X, Y, C) \ | |
2993 | ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ | |
2994 | (__v64qi)(__m512i) (Y), (int) (C), \ | |
f0515366 JJ |
2995 | (__v32hi)(__m512i) \ |
2996 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
2997 | (__mmask32)-1)) |
2998 | ||
2999 | #define _mm512_mask_dbsad_epu8(W, U, X, Y, C) \ | |
3000 | ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ | |
3001 | (__v64qi)(__m512i) (Y), (int) (C), \ | |
3002 | (__v32hi)(__m512i)(W), \ | |
3003 | (__mmask32)(U))) | |
3004 | ||
3005 | #define _mm512_maskz_dbsad_epu8(U, X, Y, C) \ | |
3006 | ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ | |
3007 | (__v64qi)(__m512i) (Y), (int) (C), \ | |
f0515366 JJ |
3008 | (__v32hi)(__m512i) \ |
3009 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3010 | (__mmask32)(U))) |
3011 | ||
3012 | #define _mm512_srli_epi16(A, B) \ | |
3013 | ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ | |
f0515366 | 3014 | (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)-1)) |
936c0fe4 AI |
3015 | |
3016 | #define _mm512_mask_srli_epi16(W, U, A, B) \ | |
3017 | ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ | |
3018 | (int)(B), (__v32hi)(__m512i)(W), (__mmask32)(U))) | |
3019 | ||
3020 | #define _mm512_maskz_srli_epi16(U, A, B) \ | |
3021 | ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ | |
f0515366 | 3022 | (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)(U))) |
936c0fe4 AI |
3023 | |
3024 | #define _mm512_slli_epi16(X, C) \ | |
3025 | ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), (int)(C),\ | |
f0515366 | 3026 | (__v32hi)(__m512i)_mm512_setzero_si512 (), \ |
936c0fe4 AI |
3027 | (__mmask32)-1)) |
3028 | ||
3029 | #define _mm512_mask_slli_epi16(W, U, X, C) \ | |
3030 | ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), (int)(C),\ | |
3031 | (__v32hi)(__m512i)(W),\ | |
3032 | (__mmask32)(U))) | |
3033 | ||
3034 | #define _mm512_maskz_slli_epi16(U, X, C) \ | |
3035 | ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), (int)(C),\ | |
f0515366 | 3036 | (__v32hi)(__m512i)_mm512_setzero_si512 (), \ |
936c0fe4 AI |
3037 | (__mmask32)(U))) |
3038 | ||
3039 | #define _mm512_shufflehi_epi16(A, B) \ | |
3040 | ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3041 | (__v32hi)(__m512i) \ |
3042 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3043 | (__mmask32)-1)) |
3044 | ||
3045 | #define _mm512_mask_shufflehi_epi16(W, U, A, B) \ | |
3046 | ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
3047 | (__v32hi)(__m512i)(W), \ | |
3048 | (__mmask32)(U))) | |
3049 | ||
3050 | #define _mm512_maskz_shufflehi_epi16(U, A, B) \ | |
3051 | ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3052 | (__v32hi)(__m512i) \ |
3053 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3054 | (__mmask32)(U))) |
3055 | ||
3056 | #define _mm512_shufflelo_epi16(A, B) \ | |
3057 | ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3058 | (__v32hi)(__m512i) \ |
3059 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3060 | (__mmask32)-1)) |
3061 | ||
3062 | #define _mm512_mask_shufflelo_epi16(W, U, A, B) \ | |
3063 | ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
3064 | (__v32hi)(__m512i)(W), \ | |
3065 | (__mmask32)(U))) | |
3066 | ||
3067 | #define _mm512_maskz_shufflelo_epi16(U, A, B) \ | |
3068 | ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3069 | (__v32hi)(__m512i) \ |
3070 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3071 | (__mmask32)(U))) |
3072 | ||
3073 | #define _mm512_srai_epi16(A, B) \ | |
3074 | ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ | |
f0515366 | 3075 | (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)-1)) |
936c0fe4 AI |
3076 | |
3077 | #define _mm512_mask_srai_epi16(W, U, A, B) \ | |
3078 | ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ | |
3079 | (int)(B), (__v32hi)(__m512i)(W), (__mmask32)(U))) | |
3080 | ||
3081 | #define _mm512_maskz_srai_epi16(U, A, B) \ | |
3082 | ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ | |
f0515366 | 3083 | (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)(U))) |
936c0fe4 AI |
3084 | |
3085 | #define _mm512_mask_blend_epi16(__U, __A, __W) \ | |
3086 | ((__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) (__A), \ | |
3087 | (__v32hi) (__W), \ | |
3088 | (__mmask32) (__U))) | |
3089 | ||
3090 | #define _mm512_mask_blend_epi8(__U, __A, __W) \ | |
3091 | ((__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) (__A), \ | |
3092 | (__v64qi) (__W), \ | |
3093 | (__mmask64) (__U))) | |
3094 | ||
3095 | #define _mm512_cmp_epi16_mask(X, Y, P) \ | |
3096 | ((__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3097 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3098 | (__mmask32)(-1))) | |
3099 | ||
3100 | #define _mm512_cmp_epi8_mask(X, Y, P) \ | |
3101 | ((__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3102 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3103 | (__mmask64)(-1))) | |
3104 | ||
3105 | #define _mm512_cmp_epu16_mask(X, Y, P) \ | |
3106 | ((__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3107 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3108 | (__mmask32)(-1))) | |
3109 | ||
3110 | #define _mm512_cmp_epu8_mask(X, Y, P) \ | |
3111 | ((__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3112 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3113 | (__mmask64)(-1))) | |
3114 | ||
3115 | #define _mm512_mask_cmp_epi16_mask(M, X, Y, P) \ | |
3116 | ((__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3117 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3118 | (__mmask32)(M))) | |
3119 | ||
3120 | #define _mm512_mask_cmp_epi8_mask(M, X, Y, P) \ | |
3121 | ((__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3122 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3123 | (__mmask64)(M))) | |
3124 | ||
3125 | #define _mm512_mask_cmp_epu16_mask(M, X, Y, P) \ | |
3126 | ((__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3127 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3128 | (__mmask32)(M))) | |
3129 | ||
3130 | #define _mm512_mask_cmp_epu8_mask(M, X, Y, P) \ | |
3131 | ((__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3132 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3133 | (__mmask64)(M))) | |
3134 | ||
3135 | #define _mm512_bslli_epi128(A, N) \ | |
3136 | ((__m512i)__builtin_ia32_pslldq512 ((__m512i)(A), (int)(N) * 8)) | |
3137 | ||
3138 | #define _mm512_bsrli_epi128(A, N) \ | |
3139 | ((__m512i)__builtin_ia32_psrldq512 ((__m512i)(A), (int)(N) * 8)) | |
3140 | ||
3141 | #endif | |
3142 | ||
3143 | #ifdef __DISABLE_AVX512BW__ | |
3144 | #undef __DISABLE_AVX512BW__ | |
3145 | #pragma GCC pop_options | |
3146 | #endif /* __DISABLE_AVX512BW__ */ | |
3147 | ||
3148 | #endif /* _AVX512BWINTRIN_H_INCLUDED */ |