]>
Commit | Line | Data |
---|---|---|
a945c346 | 1 | /* Copyright (C) 2014-2024 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 | ||
fd514717 | 31 | #if !defined (__AVX512BW__) || defined (__EVEX512__) |
936c0fe4 | 32 | #pragma GCC push_options |
fd514717 | 33 | #pragma GCC target("avx512bw,no-evex512") |
936c0fe4 AI |
34 | #define __DISABLE_AVX512BW__ |
35 | #endif /* __AVX512BW__ */ | |
36 | ||
d3545378 HJ |
37 | typedef unsigned long long __mmask64; |
38 | ||
e9529ff3 HJ |
39 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
40 | _mm_avx512_set_epi32 (int __q3, int __q2, int __q1, int __q0) | |
41 | { | |
42 | return __extension__ (__m128i)(__v4si){ __q0, __q1, __q2, __q3 }; | |
43 | } | |
44 | ||
45 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) | |
46 | _mm_avx512_set_epi16 (short __q7, short __q6, short __q5, short __q4, | |
47 | short __q3, short __q2, short __q1, short __q0) | |
48 | { | |
49 | return __extension__ (__m128i)(__v8hi){ | |
50 | __q0, __q1, __q2, __q3, __q4, __q5, __q6, __q7 }; | |
51 | } | |
52 | ||
53 | extern __inline __m128i __attribute__((__gnu_inline__, __always_inline__, __artificial__)) | |
54 | _mm_avx512_set_epi8 (char __q15, char __q14, char __q13, char __q12, | |
55 | char __q11, char __q10, char __q09, char __q08, | |
56 | char __q07, char __q06, char __q05, char __q04, | |
57 | char __q03, char __q02, char __q01, char __q00) | |
58 | { | |
59 | return __extension__ (__m128i)(__v16qi){ | |
60 | __q00, __q01, __q02, __q03, __q04, __q05, __q06, __q07, | |
61 | __q08, __q09, __q10, __q11, __q12, __q13, __q14, __q15 | |
62 | }; | |
63 | } | |
64 | ||
dea06111 AS |
65 | extern __inline unsigned char |
66 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
67 | _ktest_mask32_u8 (__mmask32 __A, __mmask32 __B, unsigned char *__CF) | |
68 | { | |
69 | *__CF = (unsigned char) __builtin_ia32_ktestcsi (__A, __B); | |
70 | return (unsigned char) __builtin_ia32_ktestzsi (__A, __B); | |
71 | } | |
72 | ||
73 | extern __inline unsigned char | |
74 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 75 | _ktestz_mask32_u8 (__mmask32 __A, __mmask32 __B) |
dea06111 | 76 | { |
03a85048 | 77 | return (unsigned char) __builtin_ia32_ktestzsi (__A, __B); |
dea06111 AS |
78 | } |
79 | ||
80 | extern __inline unsigned char | |
81 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 82 | _ktestc_mask32_u8 (__mmask32 __A, __mmask32 __B) |
dea06111 | 83 | { |
03a85048 | 84 | return (unsigned char) __builtin_ia32_ktestcsi (__A, __B); |
dea06111 AS |
85 | } |
86 | ||
87 | extern __inline unsigned char | |
88 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 89 | _kortest_mask32_u8 (__mmask32 __A, __mmask32 __B, unsigned char *__CF) |
dea06111 | 90 | { |
03a85048 HJ |
91 | *__CF = (unsigned char) __builtin_ia32_kortestcsi (__A, __B); |
92 | return (unsigned char) __builtin_ia32_kortestzsi (__A, __B); | |
dea06111 AS |
93 | } |
94 | ||
95 | extern __inline unsigned char | |
96 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 97 | _kortestz_mask32_u8 (__mmask32 __A, __mmask32 __B) |
dea06111 | 98 | { |
03a85048 | 99 | return (unsigned char) __builtin_ia32_kortestzsi (__A, __B); |
dea06111 AS |
100 | } |
101 | ||
102 | extern __inline unsigned char | |
103 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 104 | _kortestc_mask32_u8 (__mmask32 __A, __mmask32 __B) |
dea06111 | 105 | { |
03a85048 | 106 | return (unsigned char) __builtin_ia32_kortestcsi (__A, __B); |
dea06111 AS |
107 | } |
108 | ||
03a85048 | 109 | extern __inline __mmask32 |
dea06111 | 110 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 111 | _kadd_mask32 (__mmask32 __A, __mmask32 __B) |
dea06111 | 112 | { |
03a85048 | 113 | return (__mmask32) __builtin_ia32_kaddsi ((__mmask32) __A, (__mmask32) __B); |
dea06111 AS |
114 | } |
115 | ||
03a85048 | 116 | extern __inline unsigned int |
dea06111 | 117 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 118 | _cvtmask32_u32 (__mmask32 __A) |
dea06111 | 119 | { |
03a85048 | 120 | return (unsigned int) __builtin_ia32_kmovd ((__mmask32) __A); |
dea06111 AS |
121 | } |
122 | ||
03a85048 | 123 | extern __inline __mmask32 |
dea06111 | 124 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 125 | _cvtu32_mask32 (unsigned int __A) |
dea06111 | 126 | { |
03a85048 | 127 | return (__mmask32) __builtin_ia32_kmovd ((__mmask32) __A); |
dea06111 AS |
128 | } |
129 | ||
03a85048 | 130 | extern __inline __mmask32 |
dea06111 | 131 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 132 | _load_mask32 (__mmask32 *__A) |
dea06111 | 133 | { |
03a85048 | 134 | return (__mmask32) __builtin_ia32_kmovd (*__A); |
dea06111 AS |
135 | } |
136 | ||
03a85048 | 137 | extern __inline void |
dea06111 | 138 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 139 | _store_mask32 (__mmask32 *__A, __mmask32 __B) |
dea06111 | 140 | { |
03a85048 | 141 | *(__mmask32 *) __A = __builtin_ia32_kmovd (__B); |
dea06111 AS |
142 | } |
143 | ||
03a85048 | 144 | extern __inline __mmask32 |
dea06111 | 145 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 146 | _knot_mask32 (__mmask32 __A) |
dea06111 | 147 | { |
03a85048 | 148 | return (__mmask32) __builtin_ia32_knotsi ((__mmask32) __A); |
dea06111 AS |
149 | } |
150 | ||
75d727d2 AS |
151 | extern __inline __mmask32 |
152 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 153 | _kor_mask32 (__mmask32 __A, __mmask32 __B) |
75d727d2 | 154 | { |
03a85048 | 155 | return (__mmask32) __builtin_ia32_korsi ((__mmask32) __A, (__mmask32) __B); |
75d727d2 AS |
156 | } |
157 | ||
03a85048 | 158 | extern __inline __mmask32 |
75d727d2 | 159 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 160 | _kxnor_mask32 (__mmask32 __A, __mmask32 __B) |
75d727d2 | 161 | { |
03a85048 | 162 | return (__mmask32) __builtin_ia32_kxnorsi ((__mmask32) __A, (__mmask32) __B); |
75d727d2 AS |
163 | } |
164 | ||
03a85048 | 165 | extern __inline __mmask32 |
7cdb6e4c | 166 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 167 | _kxor_mask32 (__mmask32 __A, __mmask32 __B) |
7cdb6e4c | 168 | { |
03a85048 | 169 | return (__mmask32) __builtin_ia32_kxorsi ((__mmask32) __A, (__mmask32) __B); |
7cdb6e4c AS |
170 | } |
171 | ||
03a85048 | 172 | extern __inline __mmask32 |
7cdb6e4c | 173 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 174 | _kand_mask32 (__mmask32 __A, __mmask32 __B) |
7cdb6e4c | 175 | { |
03a85048 | 176 | return (__mmask32) __builtin_ia32_kandsi ((__mmask32) __A, (__mmask32) __B); |
7cdb6e4c AS |
177 | } |
178 | ||
179 | extern __inline __mmask32 | |
180 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 181 | _kandn_mask32 (__mmask32 __A, __mmask32 __B) |
7cdb6e4c | 182 | { |
03a85048 | 183 | return (__mmask32) __builtin_ia32_kandnsi ((__mmask32) __A, (__mmask32) __B); |
7cdb6e4c AS |
184 | } |
185 | ||
03a85048 | 186 | extern __inline __mmask32 |
7cdb6e4c | 187 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 188 | _mm512_kunpackw (__mmask32 __A, __mmask32 __B) |
7cdb6e4c | 189 | { |
03a85048 HJ |
190 | return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, |
191 | (__mmask32) __B); | |
7cdb6e4c AS |
192 | } |
193 | ||
194 | extern __inline __mmask32 | |
195 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 196 | _kunpackw_mask32 (__mmask16 __A, __mmask16 __B) |
7cdb6e4c | 197 | { |
03a85048 HJ |
198 | return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, |
199 | (__mmask32) __B); | |
7cdb6e4c AS |
200 | } |
201 | ||
03a85048 HJ |
202 | #if __OPTIMIZE__ |
203 | extern __inline __mmask32 | |
7cdb6e4c | 204 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 205 | _kshiftli_mask32 (__mmask32 __A, unsigned int __B) |
7cdb6e4c | 206 | { |
03a85048 HJ |
207 | return (__mmask32) __builtin_ia32_kshiftlisi ((__mmask32) __A, |
208 | (__mmask8) __B); | |
7cdb6e4c AS |
209 | } |
210 | ||
03a85048 | 211 | extern __inline __mmask32 |
7cdb6e4c | 212 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 213 | _kshiftri_mask32 (__mmask32 __A, unsigned int __B) |
7cdb6e4c | 214 | { |
03a85048 HJ |
215 | return (__mmask32) __builtin_ia32_kshiftrisi ((__mmask32) __A, |
216 | (__mmask8) __B); | |
7cdb6e4c AS |
217 | } |
218 | ||
03a85048 HJ |
219 | #else |
220 | #define _kshiftli_mask32(X, Y) \ | |
221 | ((__mmask32) __builtin_ia32_kshiftlisi ((__mmask32)(X), (__mmask8)(Y))) | |
222 | ||
223 | #define _kshiftri_mask32(X, Y) \ | |
224 | ((__mmask32) __builtin_ia32_kshiftrisi ((__mmask32)(X), (__mmask8)(Y))) | |
225 | ||
226 | #endif | |
227 | ||
03a85048 | 228 | extern __inline unsigned char |
7cdb6e4c | 229 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 230 | _ktest_mask64_u8 (__mmask64 __A, __mmask64 __B, unsigned char *__CF) |
7cdb6e4c | 231 | { |
03a85048 HJ |
232 | *__CF = (unsigned char) __builtin_ia32_ktestcdi (__A, __B); |
233 | return (unsigned char) __builtin_ia32_ktestzdi (__A, __B); | |
7cdb6e4c AS |
234 | } |
235 | ||
03a85048 | 236 | extern __inline unsigned char |
6901ea62 | 237 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 238 | _ktestz_mask64_u8 (__mmask64 __A, __mmask64 __B) |
6901ea62 | 239 | { |
03a85048 | 240 | return (unsigned char) __builtin_ia32_ktestzdi (__A, __B); |
6901ea62 AS |
241 | } |
242 | ||
03a85048 | 243 | extern __inline unsigned char |
6901ea62 | 244 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 245 | _ktestc_mask64_u8 (__mmask64 __A, __mmask64 __B) |
6901ea62 | 246 | { |
03a85048 | 247 | return (unsigned char) __builtin_ia32_ktestcdi (__A, __B); |
6901ea62 AS |
248 | } |
249 | ||
03a85048 | 250 | extern __inline unsigned char |
6901ea62 | 251 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 252 | _kortest_mask64_u8 (__mmask64 __A, __mmask64 __B, unsigned char *__CF) |
6901ea62 | 253 | { |
03a85048 HJ |
254 | *__CF = (unsigned char) __builtin_ia32_kortestcdi (__A, __B); |
255 | return (unsigned char) __builtin_ia32_kortestzdi (__A, __B); | |
256 | } | |
257 | ||
258 | extern __inline unsigned char | |
259 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
260 | _kortestz_mask64_u8 (__mmask64 __A, __mmask64 __B) | |
261 | { | |
262 | return (unsigned char) __builtin_ia32_kortestzdi (__A, __B); | |
263 | } | |
264 | ||
265 | extern __inline unsigned char | |
266 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
267 | _kortestc_mask64_u8 (__mmask64 __A, __mmask64 __B) | |
268 | { | |
269 | return (unsigned char) __builtin_ia32_kortestcdi (__A, __B); | |
6901ea62 AS |
270 | } |
271 | ||
272 | extern __inline __mmask64 | |
273 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 274 | _kadd_mask64 (__mmask64 __A, __mmask64 __B) |
6901ea62 | 275 | { |
03a85048 | 276 | return (__mmask64) __builtin_ia32_kadddi ((__mmask64) __A, (__mmask64) __B); |
6901ea62 AS |
277 | } |
278 | ||
03a85048 | 279 | extern __inline unsigned long long |
6901ea62 | 280 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 281 | _cvtmask64_u64 (__mmask64 __A) |
6901ea62 | 282 | { |
03a85048 | 283 | return (unsigned long long) __builtin_ia32_kmovq ((__mmask64) __A); |
6901ea62 AS |
284 | } |
285 | ||
286 | extern __inline __mmask64 | |
287 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 288 | _cvtu64_mask64 (unsigned long long __A) |
6901ea62 | 289 | { |
03a85048 | 290 | return (__mmask64) __builtin_ia32_kmovq ((__mmask64) __A); |
6901ea62 AS |
291 | } |
292 | ||
03a85048 | 293 | extern __inline __mmask64 |
6901ea62 | 294 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 295 | _load_mask64 (__mmask64 *__A) |
6901ea62 | 296 | { |
03a85048 HJ |
297 | return (__mmask64) __builtin_ia32_kmovq (*(__mmask64 *) __A); |
298 | } | |
299 | ||
300 | extern __inline void | |
301 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
302 | _store_mask64 (__mmask64 *__A, __mmask64 __B) | |
303 | { | |
304 | *(__mmask64 *) __A = __builtin_ia32_kmovq (__B); | |
6901ea62 AS |
305 | } |
306 | ||
307 | extern __inline __mmask64 | |
308 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 309 | _knot_mask64 (__mmask64 __A) |
6901ea62 | 310 | { |
03a85048 | 311 | return (__mmask64) __builtin_ia32_knotdi ((__mmask64) __A); |
6901ea62 AS |
312 | } |
313 | ||
03a85048 | 314 | extern __inline __mmask64 |
6901ea62 | 315 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 316 | _kor_mask64 (__mmask64 __A, __mmask64 __B) |
6901ea62 | 317 | { |
03a85048 | 318 | return (__mmask64) __builtin_ia32_kordi ((__mmask64) __A, (__mmask64) __B); |
6901ea62 AS |
319 | } |
320 | ||
321 | extern __inline __mmask64 | |
322 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
03a85048 | 323 | _kxnor_mask64 (__mmask64 __A, __mmask64 __B) |
6901ea62 | 324 | { |
03a85048 | 325 | return (__mmask64) __builtin_ia32_kxnordi ((__mmask64) __A, (__mmask64) __B); |
6901ea62 AS |
326 | } |
327 | ||
03a85048 | 328 | extern __inline __mmask64 |
6901ea62 | 329 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) |
03a85048 | 330 | _kxor_mask64 (__mmask64 __A, __mmask64 __B) |
6901ea62 | 331 | { |
03a85048 HJ |
332 | return (__mmask64) __builtin_ia32_kxordi ((__mmask64) __A, (__mmask64) __B); |
333 | } | |
334 | ||
335 | extern __inline __mmask64 | |
336 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
337 | _kand_mask64 (__mmask64 __A, __mmask64 __B) | |
338 | { | |
339 | return (__mmask64) __builtin_ia32_kanddi ((__mmask64) __A, (__mmask64) __B); | |
6901ea62 AS |
340 | } |
341 | ||
342 | extern __inline __mmask64 | |
343 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
344 | _kandn_mask64 (__mmask64 __A, __mmask64 __B) | |
345 | { | |
346 | return (__mmask64) __builtin_ia32_kandndi ((__mmask64) __A, (__mmask64) __B); | |
347 | } | |
348 | ||
d3545378 HJ |
349 | #ifdef __DISABLE_AVX512BW__ |
350 | #undef __DISABLE_AVX512BW__ | |
351 | #pragma GCC pop_options | |
352 | #endif /* __DISABLE_AVX512BW__ */ | |
353 | ||
354 | #if !defined (__AVX512BW__) || !defined (__EVEX512__) | |
355 | #pragma GCC push_options | |
356 | #pragma GCC target("avx512bw,evex512") | |
357 | #define __DISABLE_AVX512BW_512__ | |
358 | #endif /* __AVX512BW_512__ */ | |
359 | ||
360 | /* Internal data types for implementing the intrinsics. */ | |
361 | typedef short __v32hi __attribute__ ((__vector_size__ (64))); | |
362 | typedef short __v32hi_u __attribute__ ((__vector_size__ (64), \ | |
363 | __may_alias__, __aligned__ (1))); | |
364 | typedef char __v64qi __attribute__ ((__vector_size__ (64))); | |
365 | typedef char __v64qi_u __attribute__ ((__vector_size__ (64), \ | |
366 | __may_alias__, __aligned__ (1))); | |
367 | ||
936c0fe4 AI |
368 | extern __inline __m512i |
369 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
370 | _mm512_mask_mov_epi16 (__m512i __W, __mmask32 __U, __m512i __A) | |
371 | { | |
372 | return (__m512i) __builtin_ia32_movdquhi512_mask ((__v32hi) __A, | |
373 | (__v32hi) __W, | |
374 | (__mmask32) __U); | |
375 | } | |
376 | ||
377 | extern __inline __m512i | |
378 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
379 | _mm512_maskz_mov_epi16 (__mmask32 __U, __m512i __A) | |
380 | { | |
381 | return (__m512i) __builtin_ia32_movdquhi512_mask ((__v32hi) __A, | |
382 | (__v32hi) | |
f0515366 | 383 | _mm512_setzero_si512 (), |
936c0fe4 AI |
384 | (__mmask32) __U); |
385 | } | |
386 | ||
93103603 SP |
387 | extern __inline __m512i |
388 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
389 | _mm512_loadu_epi16 (void const *__P) | |
390 | { | |
391 | return (__m512i) (*(__v32hi_u *) __P); | |
392 | } | |
393 | ||
936c0fe4 AI |
394 | extern __inline __m512i |
395 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
396 | _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P) | |
397 | { | |
fc9cf6da | 398 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const short *) __P, |
936c0fe4 AI |
399 | (__v32hi) __W, |
400 | (__mmask32) __U); | |
401 | } | |
402 | ||
403 | extern __inline __m512i | |
404 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
405 | _mm512_maskz_loadu_epi16 (__mmask32 __U, void const *__P) | |
406 | { | |
fc9cf6da | 407 | return (__m512i) __builtin_ia32_loaddquhi512_mask ((const short *) __P, |
936c0fe4 | 408 | (__v32hi) |
f0515366 | 409 | _mm512_setzero_si512 (), |
936c0fe4 AI |
410 | (__mmask32) __U); |
411 | } | |
412 | ||
93103603 SP |
413 | extern __inline void |
414 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
415 | _mm512_storeu_epi16 (void *__P, __m512i __A) | |
416 | { | |
417 | *(__v32hi_u *) __P = (__v32hi_u) __A; | |
418 | } | |
419 | ||
936c0fe4 AI |
420 | extern __inline void |
421 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
422 | _mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A) | |
423 | { | |
fc9cf6da | 424 | __builtin_ia32_storedquhi512_mask ((short *) __P, |
936c0fe4 AI |
425 | (__v32hi) __A, |
426 | (__mmask32) __U); | |
427 | } | |
428 | ||
429 | extern __inline __m512i | |
430 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
431 | _mm512_mask_mov_epi8 (__m512i __W, __mmask64 __U, __m512i __A) | |
432 | { | |
433 | return (__m512i) __builtin_ia32_movdquqi512_mask ((__v64qi) __A, | |
434 | (__v64qi) __W, | |
435 | (__mmask64) __U); | |
436 | } | |
437 | ||
438 | extern __inline __m512i | |
439 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
440 | _mm512_maskz_mov_epi8 (__mmask64 __U, __m512i __A) | |
441 | { | |
442 | return (__m512i) __builtin_ia32_movdquqi512_mask ((__v64qi) __A, | |
443 | (__v64qi) | |
f0515366 | 444 | _mm512_setzero_si512 (), |
936c0fe4 AI |
445 | (__mmask64) __U); |
446 | } | |
447 | ||
936c0fe4 AI |
448 | extern __inline __mmask64 |
449 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
450 | _mm512_kunpackd (__mmask64 __A, __mmask64 __B) | |
451 | { | |
452 | return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, | |
453 | (__mmask64) __B); | |
454 | } | |
455 | ||
6901ea62 AS |
456 | extern __inline __mmask64 |
457 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
458 | _kunpackd_mask64 (__mmask32 __A, __mmask32 __B) | |
459 | { | |
460 | return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, | |
461 | (__mmask64) __B); | |
462 | } | |
463 | ||
93103603 SP |
464 | extern __inline __m512i |
465 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
466 | _mm512_loadu_epi8 (void const *__P) | |
467 | { | |
468 | return (__m512i) (*(__v64qi_u *) __P); | |
469 | } | |
470 | ||
936c0fe4 AI |
471 | extern __inline __m512i |
472 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
473 | _mm512_mask_loadu_epi8 (__m512i __W, __mmask64 __U, void const *__P) | |
474 | { | |
fc9cf6da | 475 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const char *) __P, |
936c0fe4 AI |
476 | (__v64qi) __W, |
477 | (__mmask64) __U); | |
478 | } | |
479 | ||
480 | extern __inline __m512i | |
481 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
482 | _mm512_maskz_loadu_epi8 (__mmask64 __U, void const *__P) | |
483 | { | |
fc9cf6da | 484 | return (__m512i) __builtin_ia32_loaddquqi512_mask ((const char *) __P, |
936c0fe4 | 485 | (__v64qi) |
f0515366 | 486 | _mm512_setzero_si512 (), |
936c0fe4 AI |
487 | (__mmask64) __U); |
488 | } | |
489 | ||
93103603 SP |
490 | extern __inline void |
491 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
492 | _mm512_storeu_epi8 (void *__P, __m512i __A) | |
493 | { | |
494 | *(__v64qi_u *) __P = (__v64qi_u) __A; | |
495 | } | |
496 | ||
936c0fe4 AI |
497 | extern __inline void |
498 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
499 | _mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A) | |
500 | { | |
fc9cf6da | 501 | __builtin_ia32_storedquqi512_mask ((char *) __P, |
936c0fe4 AI |
502 | (__v64qi) __A, |
503 | (__mmask64) __U); | |
504 | } | |
505 | ||
506 | extern __inline __m512i | |
507 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
508 | _mm512_sad_epu8 (__m512i __A, __m512i __B) | |
509 | { | |
510 | return (__m512i) __builtin_ia32_psadbw512 ((__v64qi) __A, | |
511 | (__v64qi) __B); | |
512 | } | |
513 | ||
514 | extern __inline __m256i | |
515 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
516 | _mm512_cvtepi16_epi8 (__m512i __A) | |
517 | { | |
518 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, | |
519 | (__v32qi) _mm256_undefined_si256(), | |
520 | (__mmask32) -1); | |
521 | } | |
522 | ||
c46f9051 JK |
523 | extern __inline void |
524 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
525 | _mm512_mask_cvtepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) | |
526 | { | |
527 | __builtin_ia32_pmovwb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); | |
528 | } | |
529 | ||
936c0fe4 AI |
530 | extern __inline __m256i |
531 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
532 | _mm512_mask_cvtepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) | |
533 | { | |
534 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, | |
535 | (__v32qi) __O, __M); | |
536 | } | |
537 | ||
538 | extern __inline __m256i | |
539 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
540 | _mm512_maskz_cvtepi16_epi8 (__mmask32 __M, __m512i __A) | |
541 | { | |
542 | return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, | |
543 | (__v32qi) | |
544 | _mm256_setzero_si256 (), | |
545 | __M); | |
546 | } | |
547 | ||
548 | extern __inline __m256i | |
549 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
550 | _mm512_cvtsepi16_epi8 (__m512i __A) | |
551 | { | |
552 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, | |
553 | (__v32qi)_mm256_undefined_si256(), | |
554 | (__mmask32) -1); | |
555 | } | |
556 | ||
c46f9051 JK |
557 | extern __inline void |
558 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
559 | _mm512_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) | |
560 | { | |
561 | __builtin_ia32_pmovswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); | |
562 | } | |
563 | ||
936c0fe4 AI |
564 | extern __inline __m256i |
565 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
566 | _mm512_mask_cvtsepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) | |
567 | { | |
568 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, | |
569 | (__v32qi)__O, | |
570 | __M); | |
571 | } | |
572 | ||
573 | extern __inline __m256i | |
574 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
575 | _mm512_maskz_cvtsepi16_epi8 (__mmask32 __M, __m512i __A) | |
576 | { | |
577 | return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, | |
578 | (__v32qi) | |
579 | _mm256_setzero_si256 (), | |
580 | __M); | |
581 | } | |
582 | ||
583 | extern __inline __m256i | |
584 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
585 | _mm512_cvtusepi16_epi8 (__m512i __A) | |
586 | { | |
587 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, | |
588 | (__v32qi)_mm256_undefined_si256(), | |
589 | (__mmask32) -1); | |
590 | } | |
591 | ||
592 | extern __inline __m256i | |
593 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
594 | _mm512_mask_cvtusepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) | |
595 | { | |
596 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, | |
597 | (__v32qi) __O, | |
598 | __M); | |
599 | } | |
600 | ||
c46f9051 JK |
601 | extern __inline void |
602 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
603 | _mm512_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) | |
604 | { | |
605 | __builtin_ia32_pmovuswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); | |
606 | } | |
607 | ||
936c0fe4 AI |
608 | extern __inline __m256i |
609 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
610 | _mm512_maskz_cvtusepi16_epi8 (__mmask32 __M, __m512i __A) | |
611 | { | |
612 | return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, | |
613 | (__v32qi) | |
614 | _mm256_setzero_si256 (), | |
615 | __M); | |
616 | } | |
617 | ||
618 | extern __inline __m512i | |
619 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
620 | _mm512_broadcastb_epi8 (__m128i __A) | |
621 | { | |
622 | return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, | |
4271e5cb UB |
623 | (__v64qi)_mm512_undefined_epi32(), |
624 | (__mmask64) -1); | |
936c0fe4 AI |
625 | } |
626 | ||
627 | extern __inline __m512i | |
628 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
629 | _mm512_mask_broadcastb_epi8 (__m512i __O, __mmask64 __M, __m128i __A) | |
630 | { | |
631 | return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, | |
632 | (__v64qi) __O, | |
633 | __M); | |
634 | } | |
635 | ||
636 | extern __inline __m512i | |
637 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
638 | _mm512_maskz_broadcastb_epi8 (__mmask64 __M, __m128i __A) | |
639 | { | |
640 | return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, | |
641 | (__v64qi) | |
f0515366 | 642 | _mm512_setzero_si512 (), |
936c0fe4 AI |
643 | __M); |
644 | } | |
645 | ||
646 | extern __inline __m512i | |
647 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
648 | _mm512_mask_set1_epi8 (__m512i __O, __mmask64 __M, char __A) | |
649 | { | |
650 | return (__m512i) __builtin_ia32_pbroadcastb512_gpr_mask (__A, | |
651 | (__v64qi) __O, | |
652 | __M); | |
653 | } | |
654 | ||
655 | extern __inline __m512i | |
656 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
657 | _mm512_maskz_set1_epi8 (__mmask64 __M, char __A) | |
658 | { | |
f0515366 JJ |
659 | return (__m512i) |
660 | __builtin_ia32_pbroadcastb512_gpr_mask (__A, | |
661 | (__v64qi) | |
662 | _mm512_setzero_si512 (), | |
663 | __M); | |
936c0fe4 AI |
664 | } |
665 | ||
666 | extern __inline __m512i | |
667 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
668 | _mm512_broadcastw_epi16 (__m128i __A) | |
669 | { | |
670 | return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, | |
4271e5cb UB |
671 | (__v32hi)_mm512_undefined_epi32(), |
672 | (__mmask32) -1); | |
936c0fe4 AI |
673 | } |
674 | ||
675 | extern __inline __m512i | |
676 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
677 | _mm512_mask_broadcastw_epi16 (__m512i __O, __mmask32 __M, __m128i __A) | |
678 | { | |
679 | return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, | |
680 | (__v32hi) __O, | |
681 | __M); | |
682 | } | |
683 | ||
684 | extern __inline __m512i | |
685 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
686 | _mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A) | |
687 | { | |
688 | return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, | |
689 | (__v32hi) | |
f0515366 | 690 | _mm512_setzero_si512 (), |
936c0fe4 AI |
691 | __M); |
692 | } | |
693 | ||
694 | extern __inline __m512i | |
695 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
696 | _mm512_mask_set1_epi16 (__m512i __O, __mmask32 __M, short __A) | |
697 | { | |
698 | return (__m512i) __builtin_ia32_pbroadcastw512_gpr_mask (__A, | |
699 | (__v32hi) __O, | |
700 | __M); | |
701 | } | |
702 | ||
703 | extern __inline __m512i | |
704 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
705 | _mm512_maskz_set1_epi16 (__mmask32 __M, short __A) | |
706 | { | |
f0515366 JJ |
707 | return (__m512i) |
708 | __builtin_ia32_pbroadcastw512_gpr_mask (__A, | |
709 | (__v32hi) | |
710 | _mm512_setzero_si512 (), | |
711 | __M); | |
936c0fe4 AI |
712 | } |
713 | ||
714 | extern __inline __m512i | |
715 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
716 | _mm512_mulhrs_epi16 (__m512i __A, __m512i __B) | |
717 | { | |
718 | return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, | |
719 | (__v32hi) __B, | |
720 | (__v32hi) | |
f0515366 | 721 | _mm512_setzero_si512 (), |
936c0fe4 AI |
722 | (__mmask32) -1); |
723 | } | |
724 | ||
725 | extern __inline __m512i | |
726 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
727 | _mm512_mask_mulhrs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
728 | __m512i __B) | |
729 | { | |
730 | return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, | |
731 | (__v32hi) __B, | |
732 | (__v32hi) __W, | |
733 | (__mmask32) __U); | |
734 | } | |
735 | ||
736 | extern __inline __m512i | |
737 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
738 | _mm512_maskz_mulhrs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
739 | { | |
740 | return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, | |
741 | (__v32hi) __B, | |
742 | (__v32hi) | |
f0515366 | 743 | _mm512_setzero_si512 (), |
936c0fe4 AI |
744 | (__mmask32) __U); |
745 | } | |
746 | ||
747 | extern __inline __m512i | |
748 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
749 | _mm512_mulhi_epi16 (__m512i __A, __m512i __B) | |
750 | { | |
751 | return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, | |
752 | (__v32hi) __B, | |
753 | (__v32hi) | |
f0515366 | 754 | _mm512_setzero_si512 (), |
936c0fe4 AI |
755 | (__mmask32) -1); |
756 | } | |
757 | ||
758 | extern __inline __m512i | |
759 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
760 | _mm512_mask_mulhi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
761 | __m512i __B) | |
762 | { | |
763 | return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, | |
764 | (__v32hi) __B, | |
765 | (__v32hi) __W, | |
766 | (__mmask32) __U); | |
767 | } | |
768 | ||
769 | extern __inline __m512i | |
770 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
771 | _mm512_maskz_mulhi_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
772 | { | |
773 | return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, | |
774 | (__v32hi) __B, | |
775 | (__v32hi) | |
f0515366 | 776 | _mm512_setzero_si512 (), |
936c0fe4 AI |
777 | (__mmask32) __U); |
778 | } | |
779 | ||
780 | extern __inline __m512i | |
781 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
782 | _mm512_mulhi_epu16 (__m512i __A, __m512i __B) | |
783 | { | |
784 | return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, | |
785 | (__v32hi) __B, | |
786 | (__v32hi) | |
f0515366 | 787 | _mm512_setzero_si512 (), |
936c0fe4 AI |
788 | (__mmask32) -1); |
789 | } | |
790 | ||
791 | extern __inline __m512i | |
792 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
793 | _mm512_mask_mulhi_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
794 | __m512i __B) | |
795 | { | |
796 | return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, | |
797 | (__v32hi) __B, | |
798 | (__v32hi) __W, | |
799 | (__mmask32) __U); | |
800 | } | |
801 | ||
802 | extern __inline __m512i | |
803 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
804 | _mm512_maskz_mulhi_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
805 | { | |
806 | return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, | |
807 | (__v32hi) __B, | |
808 | (__v32hi) | |
f0515366 | 809 | _mm512_setzero_si512 (), |
936c0fe4 AI |
810 | (__mmask32) __U); |
811 | } | |
812 | ||
813 | extern __inline __m512i | |
814 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
815 | _mm512_mullo_epi16 (__m512i __A, __m512i __B) | |
816 | { | |
2069d6fc | 817 | return (__m512i) ((__v32hu) __A * (__v32hu) __B); |
936c0fe4 AI |
818 | } |
819 | ||
820 | extern __inline __m512i | |
821 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
822 | _mm512_mask_mullo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
823 | __m512i __B) | |
824 | { | |
825 | return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A, | |
826 | (__v32hi) __B, | |
827 | (__v32hi) __W, | |
828 | (__mmask32) __U); | |
829 | } | |
830 | ||
831 | extern __inline __m512i | |
832 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
833 | _mm512_maskz_mullo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
834 | { | |
835 | return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A, | |
836 | (__v32hi) __B, | |
837 | (__v32hi) | |
f0515366 | 838 | _mm512_setzero_si512 (), |
936c0fe4 AI |
839 | (__mmask32) __U); |
840 | } | |
841 | ||
842 | extern __inline __m512i | |
843 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
844 | _mm512_cvtepi8_epi16 (__m256i __A) | |
845 | { | |
846 | return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, | |
847 | (__v32hi) | |
f0515366 | 848 | _mm512_setzero_si512 (), |
936c0fe4 AI |
849 | (__mmask32) -1); |
850 | } | |
851 | ||
852 | extern __inline __m512i | |
853 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
854 | _mm512_mask_cvtepi8_epi16 (__m512i __W, __mmask32 __U, __m256i __A) | |
855 | { | |
856 | return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, | |
857 | (__v32hi) __W, | |
858 | (__mmask32) __U); | |
859 | } | |
860 | ||
861 | extern __inline __m512i | |
862 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
863 | _mm512_maskz_cvtepi8_epi16 (__mmask32 __U, __m256i __A) | |
864 | { | |
865 | return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, | |
866 | (__v32hi) | |
f0515366 | 867 | _mm512_setzero_si512 (), |
936c0fe4 AI |
868 | (__mmask32) __U); |
869 | } | |
870 | ||
871 | extern __inline __m512i | |
872 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
873 | _mm512_cvtepu8_epi16 (__m256i __A) | |
874 | { | |
875 | return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, | |
876 | (__v32hi) | |
f0515366 | 877 | _mm512_setzero_si512 (), |
936c0fe4 AI |
878 | (__mmask32) -1); |
879 | } | |
880 | ||
881 | extern __inline __m512i | |
882 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
883 | _mm512_mask_cvtepu8_epi16 (__m512i __W, __mmask32 __U, __m256i __A) | |
884 | { | |
885 | return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, | |
886 | (__v32hi) __W, | |
887 | (__mmask32) __U); | |
888 | } | |
889 | ||
890 | extern __inline __m512i | |
891 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
892 | _mm512_maskz_cvtepu8_epi16 (__mmask32 __U, __m256i __A) | |
893 | { | |
894 | return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, | |
895 | (__v32hi) | |
f0515366 | 896 | _mm512_setzero_si512 (), |
936c0fe4 AI |
897 | (__mmask32) __U); |
898 | } | |
899 | ||
900 | extern __inline __m512i | |
901 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
902 | _mm512_permutexvar_epi16 (__m512i __A, __m512i __B) | |
903 | { | |
904 | return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, | |
905 | (__v32hi) __A, | |
906 | (__v32hi) | |
f0515366 | 907 | _mm512_setzero_si512 (), |
936c0fe4 AI |
908 | (__mmask32) -1); |
909 | } | |
910 | ||
911 | extern __inline __m512i | |
912 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
913 | _mm512_maskz_permutexvar_epi16 (__mmask32 __M, __m512i __A, | |
914 | __m512i __B) | |
915 | { | |
916 | return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, | |
917 | (__v32hi) __A, | |
918 | (__v32hi) | |
f0515366 | 919 | _mm512_setzero_si512 (), |
936c0fe4 AI |
920 | (__mmask32) __M); |
921 | } | |
922 | ||
923 | extern __inline __m512i | |
924 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
925 | _mm512_mask_permutexvar_epi16 (__m512i __W, __mmask32 __M, __m512i __A, | |
926 | __m512i __B) | |
927 | { | |
928 | return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, | |
929 | (__v32hi) __A, | |
930 | (__v32hi) __W, | |
931 | (__mmask32) __M); | |
932 | } | |
933 | ||
934 | extern __inline __m512i | |
935 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
936 | _mm512_permutex2var_epi16 (__m512i __A, __m512i __I, __m512i __B) | |
937 | { | |
938 | return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I | |
939 | /* idx */ , | |
940 | (__v32hi) __A, | |
941 | (__v32hi) __B, | |
281f7120 | 942 | (__mmask32) -1); |
936c0fe4 AI |
943 | } |
944 | ||
945 | extern __inline __m512i | |
946 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
947 | _mm512_mask_permutex2var_epi16 (__m512i __A, __mmask32 __U, | |
948 | __m512i __I, __m512i __B) | |
949 | { | |
950 | return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I | |
951 | /* idx */ , | |
952 | (__v32hi) __A, | |
953 | (__v32hi) __B, | |
954 | (__mmask32) | |
955 | __U); | |
956 | } | |
957 | ||
958 | extern __inline __m512i | |
959 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
960 | _mm512_mask2_permutex2var_epi16 (__m512i __A, __m512i __I, | |
961 | __mmask32 __U, __m512i __B) | |
962 | { | |
963 | return (__m512i) __builtin_ia32_vpermi2varhi512_mask ((__v32hi) __A, | |
964 | (__v32hi) __I | |
965 | /* idx */ , | |
966 | (__v32hi) __B, | |
967 | (__mmask32) | |
968 | __U); | |
969 | } | |
970 | ||
971 | extern __inline __m512i | |
972 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
973 | _mm512_maskz_permutex2var_epi16 (__mmask32 __U, __m512i __A, | |
974 | __m512i __I, __m512i __B) | |
975 | { | |
976 | return (__m512i) __builtin_ia32_vpermt2varhi512_maskz ((__v32hi) __I | |
977 | /* idx */ , | |
978 | (__v32hi) __A, | |
979 | (__v32hi) __B, | |
980 | (__mmask32) | |
981 | __U); | |
982 | } | |
983 | ||
984 | extern __inline __m512i | |
985 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
986 | _mm512_avg_epu8 (__m512i __A, __m512i __B) | |
987 | { | |
988 | return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, | |
989 | (__v64qi) __B, | |
990 | (__v64qi) | |
f0515366 | 991 | _mm512_setzero_si512 (), |
936c0fe4 AI |
992 | (__mmask64) -1); |
993 | } | |
994 | ||
995 | extern __inline __m512i | |
996 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
997 | _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A, | |
998 | __m512i __B) | |
999 | { | |
1000 | return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, | |
1001 | (__v64qi) __B, | |
1002 | (__v64qi) __W, | |
1003 | (__mmask64) __U); | |
1004 | } | |
1005 | ||
1006 | extern __inline __m512i | |
1007 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1008 | _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1009 | { | |
1010 | return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, | |
1011 | (__v64qi) __B, | |
1012 | (__v64qi) | |
f0515366 | 1013 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1014 | (__mmask64) __U); |
1015 | } | |
1016 | ||
1017 | extern __inline __m512i | |
1018 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1019 | _mm512_add_epi8 (__m512i __A, __m512i __B) | |
1020 | { | |
2069d6fc | 1021 | return (__m512i) ((__v64qu) __A + (__v64qu) __B); |
936c0fe4 AI |
1022 | } |
1023 | ||
1024 | extern __inline __m512i | |
1025 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1026 | _mm512_mask_add_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1027 | __m512i __B) | |
1028 | { | |
1029 | return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A, | |
1030 | (__v64qi) __B, | |
1031 | (__v64qi) __W, | |
1032 | (__mmask64) __U); | |
1033 | } | |
1034 | ||
1035 | extern __inline __m512i | |
1036 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1037 | _mm512_maskz_add_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1038 | { | |
1039 | return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A, | |
1040 | (__v64qi) __B, | |
1041 | (__v64qi) | |
f0515366 | 1042 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1043 | (__mmask64) __U); |
1044 | } | |
1045 | ||
1046 | extern __inline __m512i | |
1047 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1048 | _mm512_sub_epi8 (__m512i __A, __m512i __B) | |
1049 | { | |
2069d6fc | 1050 | return (__m512i) ((__v64qu) __A - (__v64qu) __B); |
936c0fe4 AI |
1051 | } |
1052 | ||
1053 | extern __inline __m512i | |
1054 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1055 | _mm512_mask_sub_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1056 | __m512i __B) | |
1057 | { | |
1058 | return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A, | |
1059 | (__v64qi) __B, | |
1060 | (__v64qi) __W, | |
1061 | (__mmask64) __U); | |
1062 | } | |
1063 | ||
1064 | extern __inline __m512i | |
1065 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1066 | _mm512_maskz_sub_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1067 | { | |
1068 | return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A, | |
1069 | (__v64qi) __B, | |
1070 | (__v64qi) | |
f0515366 | 1071 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1072 | (__mmask64) __U); |
1073 | } | |
1074 | ||
1075 | extern __inline __m512i | |
1076 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1077 | _mm512_avg_epu16 (__m512i __A, __m512i __B) | |
1078 | { | |
1079 | return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, | |
1080 | (__v32hi) __B, | |
1081 | (__v32hi) | |
f0515366 | 1082 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1083 | (__mmask32) -1); |
1084 | } | |
1085 | ||
1086 | extern __inline __m512i | |
1087 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1088 | _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1089 | __m512i __B) | |
1090 | { | |
1091 | return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, | |
1092 | (__v32hi) __B, | |
1093 | (__v32hi) __W, | |
1094 | (__mmask32) __U); | |
1095 | } | |
1096 | ||
1097 | extern __inline __m512i | |
1098 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1099 | _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1100 | { | |
1101 | return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, | |
1102 | (__v32hi) __B, | |
1103 | (__v32hi) | |
f0515366 | 1104 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1105 | (__mmask32) __U); |
1106 | } | |
1107 | ||
1108 | extern __inline __m512i | |
1109 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1110 | _mm512_subs_epi8 (__m512i __A, __m512i __B) | |
1111 | { | |
1112 | return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, | |
1113 | (__v64qi) __B, | |
1114 | (__v64qi) | |
f0515366 | 1115 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1116 | (__mmask64) -1); |
1117 | } | |
1118 | ||
1119 | extern __inline __m512i | |
1120 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1121 | _mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1122 | __m512i __B) | |
1123 | { | |
1124 | return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, | |
1125 | (__v64qi) __B, | |
1126 | (__v64qi) __W, | |
1127 | (__mmask64) __U); | |
1128 | } | |
1129 | ||
1130 | extern __inline __m512i | |
1131 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1132 | _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1133 | { | |
1134 | return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, | |
1135 | (__v64qi) __B, | |
1136 | (__v64qi) | |
f0515366 | 1137 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1138 | (__mmask64) __U); |
1139 | } | |
1140 | ||
1141 | extern __inline __m512i | |
1142 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1143 | _mm512_subs_epu8 (__m512i __A, __m512i __B) | |
1144 | { | |
1145 | return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, | |
1146 | (__v64qi) __B, | |
1147 | (__v64qi) | |
f0515366 | 1148 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1149 | (__mmask64) -1); |
1150 | } | |
1151 | ||
1152 | extern __inline __m512i | |
1153 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1154 | _mm512_mask_subs_epu8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1155 | __m512i __B) | |
1156 | { | |
1157 | return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, | |
1158 | (__v64qi) __B, | |
1159 | (__v64qi) __W, | |
1160 | (__mmask64) __U); | |
1161 | } | |
1162 | ||
1163 | extern __inline __m512i | |
1164 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1165 | _mm512_maskz_subs_epu8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1166 | { | |
1167 | return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, | |
1168 | (__v64qi) __B, | |
1169 | (__v64qi) | |
f0515366 | 1170 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1171 | (__mmask64) __U); |
1172 | } | |
1173 | ||
1174 | extern __inline __m512i | |
1175 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1176 | _mm512_adds_epi8 (__m512i __A, __m512i __B) | |
1177 | { | |
1178 | return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, | |
1179 | (__v64qi) __B, | |
1180 | (__v64qi) | |
f0515366 | 1181 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1182 | (__mmask64) -1); |
1183 | } | |
1184 | ||
1185 | extern __inline __m512i | |
1186 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1187 | _mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1188 | __m512i __B) | |
1189 | { | |
1190 | return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, | |
1191 | (__v64qi) __B, | |
1192 | (__v64qi) __W, | |
1193 | (__mmask64) __U); | |
1194 | } | |
1195 | ||
1196 | extern __inline __m512i | |
1197 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1198 | _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1199 | { | |
1200 | return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, | |
1201 | (__v64qi) __B, | |
1202 | (__v64qi) | |
f0515366 | 1203 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1204 | (__mmask64) __U); |
1205 | } | |
1206 | ||
1207 | extern __inline __m512i | |
1208 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1209 | _mm512_adds_epu8 (__m512i __A, __m512i __B) | |
1210 | { | |
1211 | return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, | |
1212 | (__v64qi) __B, | |
1213 | (__v64qi) | |
f0515366 | 1214 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1215 | (__mmask64) -1); |
1216 | } | |
1217 | ||
1218 | extern __inline __m512i | |
1219 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1220 | _mm512_mask_adds_epu8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1221 | __m512i __B) | |
1222 | { | |
1223 | return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, | |
1224 | (__v64qi) __B, | |
1225 | (__v64qi) __W, | |
1226 | (__mmask64) __U); | |
1227 | } | |
1228 | ||
1229 | extern __inline __m512i | |
1230 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1231 | _mm512_maskz_adds_epu8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1232 | { | |
1233 | return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, | |
1234 | (__v64qi) __B, | |
1235 | (__v64qi) | |
f0515366 | 1236 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1237 | (__mmask64) __U); |
1238 | } | |
1239 | ||
1240 | extern __inline __m512i | |
1241 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1242 | _mm512_sub_epi16 (__m512i __A, __m512i __B) | |
1243 | { | |
2069d6fc | 1244 | return (__m512i) ((__v32hu) __A - (__v32hu) __B); |
936c0fe4 AI |
1245 | } |
1246 | ||
1247 | extern __inline __m512i | |
1248 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1249 | _mm512_mask_sub_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1250 | __m512i __B) | |
1251 | { | |
1252 | return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A, | |
1253 | (__v32hi) __B, | |
1254 | (__v32hi) __W, | |
1255 | (__mmask32) __U); | |
1256 | } | |
1257 | ||
1258 | extern __inline __m512i | |
1259 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1260 | _mm512_maskz_sub_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1261 | { | |
1262 | return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A, | |
1263 | (__v32hi) __B, | |
1264 | (__v32hi) | |
f0515366 | 1265 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1266 | (__mmask32) __U); |
1267 | } | |
1268 | ||
1269 | extern __inline __m512i | |
1270 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1271 | _mm512_subs_epi16 (__m512i __A, __m512i __B) | |
1272 | { | |
1273 | return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, | |
1274 | (__v32hi) __B, | |
1275 | (__v32hi) | |
f0515366 | 1276 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1277 | (__mmask32) -1); |
1278 | } | |
1279 | ||
1280 | extern __inline __m512i | |
1281 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1282 | _mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1283 | __m512i __B) | |
1284 | { | |
1285 | return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, | |
1286 | (__v32hi) __B, | |
1287 | (__v32hi) __W, | |
1288 | (__mmask32) __U); | |
1289 | } | |
1290 | ||
1291 | extern __inline __m512i | |
1292 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1293 | _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1294 | { | |
1295 | return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, | |
1296 | (__v32hi) __B, | |
1297 | (__v32hi) | |
f0515366 | 1298 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1299 | (__mmask32) __U); |
1300 | } | |
1301 | ||
1302 | extern __inline __m512i | |
1303 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1304 | _mm512_subs_epu16 (__m512i __A, __m512i __B) | |
1305 | { | |
1306 | return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, | |
1307 | (__v32hi) __B, | |
1308 | (__v32hi) | |
f0515366 | 1309 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1310 | (__mmask32) -1); |
1311 | } | |
1312 | ||
1313 | extern __inline __m512i | |
1314 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1315 | _mm512_mask_subs_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1316 | __m512i __B) | |
1317 | { | |
1318 | return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, | |
1319 | (__v32hi) __B, | |
1320 | (__v32hi) __W, | |
1321 | (__mmask32) __U); | |
1322 | } | |
1323 | ||
1324 | extern __inline __m512i | |
1325 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1326 | _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1327 | { | |
1328 | return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, | |
1329 | (__v32hi) __B, | |
1330 | (__v32hi) | |
f0515366 | 1331 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1332 | (__mmask32) __U); |
1333 | } | |
1334 | ||
1335 | extern __inline __m512i | |
1336 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1337 | _mm512_add_epi16 (__m512i __A, __m512i __B) | |
1338 | { | |
2069d6fc | 1339 | return (__m512i) ((__v32hu) __A + (__v32hu) __B); |
936c0fe4 AI |
1340 | } |
1341 | ||
1342 | extern __inline __m512i | |
1343 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1344 | _mm512_mask_add_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1345 | __m512i __B) | |
1346 | { | |
1347 | return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A, | |
1348 | (__v32hi) __B, | |
1349 | (__v32hi) __W, | |
1350 | (__mmask32) __U); | |
1351 | } | |
1352 | ||
1353 | extern __inline __m512i | |
1354 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1355 | _mm512_maskz_add_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1356 | { | |
1357 | return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A, | |
1358 | (__v32hi) __B, | |
1359 | (__v32hi) | |
f0515366 | 1360 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1361 | (__mmask32) __U); |
1362 | } | |
1363 | ||
1364 | extern __inline __m512i | |
1365 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1366 | _mm512_adds_epi16 (__m512i __A, __m512i __B) | |
1367 | { | |
1368 | return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, | |
1369 | (__v32hi) __B, | |
1370 | (__v32hi) | |
f0515366 | 1371 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1372 | (__mmask32) -1); |
1373 | } | |
1374 | ||
1375 | extern __inline __m512i | |
1376 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1377 | _mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1378 | __m512i __B) | |
1379 | { | |
1380 | return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, | |
1381 | (__v32hi) __B, | |
1382 | (__v32hi) __W, | |
1383 | (__mmask32) __U); | |
1384 | } | |
1385 | ||
1386 | extern __inline __m512i | |
1387 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1388 | _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1389 | { | |
1390 | return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, | |
1391 | (__v32hi) __B, | |
1392 | (__v32hi) | |
f0515366 | 1393 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1394 | (__mmask32) __U); |
1395 | } | |
1396 | ||
1397 | extern __inline __m512i | |
1398 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1399 | _mm512_adds_epu16 (__m512i __A, __m512i __B) | |
1400 | { | |
1401 | return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, | |
1402 | (__v32hi) __B, | |
1403 | (__v32hi) | |
f0515366 | 1404 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1405 | (__mmask32) -1); |
1406 | } | |
1407 | ||
1408 | extern __inline __m512i | |
1409 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1410 | _mm512_mask_adds_epu16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1411 | __m512i __B) | |
1412 | { | |
1413 | return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, | |
1414 | (__v32hi) __B, | |
1415 | (__v32hi) __W, | |
1416 | (__mmask32) __U); | |
1417 | } | |
1418 | ||
1419 | extern __inline __m512i | |
1420 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1421 | _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1422 | { | |
1423 | return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, | |
1424 | (__v32hi) __B, | |
1425 | (__v32hi) | |
f0515366 | 1426 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1427 | (__mmask32) __U); |
1428 | } | |
1429 | ||
1430 | extern __inline __m512i | |
1431 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1432 | _mm512_srl_epi16 (__m512i __A, __m128i __B) | |
1433 | { | |
1434 | return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, | |
1435 | (__v8hi) __B, | |
1436 | (__v32hi) | |
f0515366 | 1437 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1438 | (__mmask32) -1); |
1439 | } | |
1440 | ||
1441 | extern __inline __m512i | |
1442 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1443 | _mm512_mask_srl_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1444 | __m128i __B) | |
1445 | { | |
1446 | return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, | |
1447 | (__v8hi) __B, | |
1448 | (__v32hi) __W, | |
1449 | (__mmask32) __U); | |
1450 | } | |
1451 | ||
1452 | extern __inline __m512i | |
1453 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1454 | _mm512_maskz_srl_epi16 (__mmask32 __U, __m512i __A, __m128i __B) | |
1455 | { | |
1456 | return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, | |
1457 | (__v8hi) __B, | |
1458 | (__v32hi) | |
f0515366 | 1459 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1460 | (__mmask32) __U); |
1461 | } | |
1462 | ||
1463 | extern __inline __m512i | |
1464 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1465 | _mm512_packs_epi16 (__m512i __A, __m512i __B) | |
1466 | { | |
1467 | return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, | |
1468 | (__v32hi) __B, | |
1469 | (__v64qi) | |
f0515366 | 1470 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1471 | (__mmask64) -1); |
1472 | } | |
1473 | ||
1474 | extern __inline __m512i | |
1475 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1476 | _mm512_sll_epi16 (__m512i __A, __m128i __B) | |
1477 | { | |
1478 | return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, | |
1479 | (__v8hi) __B, | |
1480 | (__v32hi) | |
f0515366 | 1481 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1482 | (__mmask32) -1); |
1483 | } | |
1484 | ||
1485 | extern __inline __m512i | |
1486 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1487 | _mm512_mask_sll_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1488 | __m128i __B) | |
1489 | { | |
1490 | return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, | |
1491 | (__v8hi) __B, | |
1492 | (__v32hi) __W, | |
1493 | (__mmask32) __U); | |
1494 | } | |
1495 | ||
1496 | extern __inline __m512i | |
1497 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1498 | _mm512_maskz_sll_epi16 (__mmask32 __U, __m512i __A, __m128i __B) | |
1499 | { | |
1500 | return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, | |
1501 | (__v8hi) __B, | |
1502 | (__v32hi) | |
f0515366 | 1503 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1504 | (__mmask32) __U); |
1505 | } | |
1506 | ||
1507 | extern __inline __m512i | |
1508 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1509 | _mm512_maddubs_epi16 (__m512i __X, __m512i __Y) | |
1510 | { | |
1511 | return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, | |
1512 | (__v64qi) __Y, | |
1513 | (__v32hi) | |
f0515366 | 1514 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1515 | (__mmask32) -1); |
1516 | } | |
1517 | ||
1518 | extern __inline __m512i | |
1519 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1520 | _mm512_mask_maddubs_epi16 (__m512i __W, __mmask32 __U, __m512i __X, | |
1521 | __m512i __Y) | |
1522 | { | |
1523 | return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, | |
1524 | (__v64qi) __Y, | |
1525 | (__v32hi) __W, | |
1526 | (__mmask32) __U); | |
1527 | } | |
1528 | ||
1529 | extern __inline __m512i | |
1530 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1531 | _mm512_maskz_maddubs_epi16 (__mmask32 __U, __m512i __X, __m512i __Y) | |
1532 | { | |
1533 | return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, | |
1534 | (__v64qi) __Y, | |
1535 | (__v32hi) | |
f0515366 | 1536 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1537 | (__mmask32) __U); |
1538 | } | |
1539 | ||
1540 | extern __inline __m512i | |
1541 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1542 | _mm512_madd_epi16 (__m512i __A, __m512i __B) | |
1543 | { | |
1544 | return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, | |
1545 | (__v32hi) __B, | |
1546 | (__v16si) | |
1547 | _mm512_setzero_si512 (), | |
1548 | (__mmask16) -1); | |
1549 | } | |
1550 | ||
1551 | extern __inline __m512i | |
1552 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1553 | _mm512_mask_madd_epi16 (__m512i __W, __mmask16 __U, __m512i __A, | |
1554 | __m512i __B) | |
1555 | { | |
1556 | return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, | |
1557 | (__v32hi) __B, | |
1558 | (__v16si) __W, | |
1559 | (__mmask16) __U); | |
1560 | } | |
1561 | ||
1562 | extern __inline __m512i | |
1563 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1564 | _mm512_maskz_madd_epi16 (__mmask16 __U, __m512i __A, __m512i __B) | |
1565 | { | |
1566 | return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, | |
1567 | (__v32hi) __B, | |
1568 | (__v16si) | |
1569 | _mm512_setzero_si512 (), | |
1570 | (__mmask16) __U); | |
1571 | } | |
1572 | ||
1573 | extern __inline __m512i | |
1574 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1575 | _mm512_unpackhi_epi8 (__m512i __A, __m512i __B) | |
1576 | { | |
1577 | return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, | |
1578 | (__v64qi) __B, | |
1579 | (__v64qi) | |
f0515366 | 1580 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1581 | (__mmask64) -1); |
1582 | } | |
1583 | ||
1584 | extern __inline __m512i | |
1585 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1586 | _mm512_mask_unpackhi_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1587 | __m512i __B) | |
1588 | { | |
1589 | return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, | |
1590 | (__v64qi) __B, | |
1591 | (__v64qi) __W, | |
1592 | (__mmask64) __U); | |
1593 | } | |
1594 | ||
1595 | extern __inline __m512i | |
1596 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1597 | _mm512_maskz_unpackhi_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1598 | { | |
1599 | return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, | |
1600 | (__v64qi) __B, | |
1601 | (__v64qi) | |
f0515366 | 1602 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1603 | (__mmask64) __U); |
1604 | } | |
1605 | ||
1606 | extern __inline __m512i | |
1607 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1608 | _mm512_unpackhi_epi16 (__m512i __A, __m512i __B) | |
1609 | { | |
1610 | return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, | |
1611 | (__v32hi) __B, | |
1612 | (__v32hi) | |
f0515366 | 1613 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1614 | (__mmask32) -1); |
1615 | } | |
1616 | ||
1617 | extern __inline __m512i | |
1618 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1619 | _mm512_mask_unpackhi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1620 | __m512i __B) | |
1621 | { | |
1622 | return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, | |
1623 | (__v32hi) __B, | |
1624 | (__v32hi) __W, | |
1625 | (__mmask32) __U); | |
1626 | } | |
1627 | ||
1628 | extern __inline __m512i | |
1629 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1630 | _mm512_maskz_unpackhi_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1631 | { | |
1632 | return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, | |
1633 | (__v32hi) __B, | |
1634 | (__v32hi) | |
f0515366 | 1635 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1636 | (__mmask32) __U); |
1637 | } | |
1638 | ||
1639 | extern __inline __m512i | |
1640 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1641 | _mm512_unpacklo_epi8 (__m512i __A, __m512i __B) | |
1642 | { | |
1643 | return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, | |
1644 | (__v64qi) __B, | |
1645 | (__v64qi) | |
f0515366 | 1646 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1647 | (__mmask64) -1); |
1648 | } | |
1649 | ||
1650 | extern __inline __m512i | |
1651 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1652 | _mm512_mask_unpacklo_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1653 | __m512i __B) | |
1654 | { | |
1655 | return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, | |
1656 | (__v64qi) __B, | |
1657 | (__v64qi) __W, | |
1658 | (__mmask64) __U); | |
1659 | } | |
1660 | ||
1661 | extern __inline __m512i | |
1662 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1663 | _mm512_maskz_unpacklo_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1664 | { | |
1665 | return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, | |
1666 | (__v64qi) __B, | |
1667 | (__v64qi) | |
f0515366 | 1668 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1669 | (__mmask64) __U); |
1670 | } | |
1671 | ||
1672 | extern __inline __m512i | |
1673 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1674 | _mm512_unpacklo_epi16 (__m512i __A, __m512i __B) | |
1675 | { | |
1676 | return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, | |
1677 | (__v32hi) __B, | |
1678 | (__v32hi) | |
f0515366 | 1679 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1680 | (__mmask32) -1); |
1681 | } | |
1682 | ||
1683 | extern __inline __m512i | |
1684 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1685 | _mm512_mask_unpacklo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
1686 | __m512i __B) | |
1687 | { | |
1688 | return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, | |
1689 | (__v32hi) __B, | |
1690 | (__v32hi) __W, | |
1691 | (__mmask32) __U); | |
1692 | } | |
1693 | ||
1694 | extern __inline __m512i | |
1695 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1696 | _mm512_maskz_unpacklo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
1697 | { | |
1698 | return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, | |
1699 | (__v32hi) __B, | |
1700 | (__v32hi) | |
f0515366 | 1701 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1702 | (__mmask32) __U); |
1703 | } | |
1704 | ||
eee5d6f5 AI |
1705 | extern __inline __mmask64 |
1706 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1707 | _mm512_cmpeq_epu8_mask (__m512i __A, __m512i __B) | |
1708 | { | |
1709 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1710 | (__v64qi) __B, 0, | |
1711 | (__mmask64) -1); | |
1712 | } | |
1713 | ||
936c0fe4 AI |
1714 | extern __inline __mmask64 |
1715 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1716 | _mm512_cmpeq_epi8_mask (__m512i __A, __m512i __B) | |
1717 | { | |
1718 | return (__mmask64) __builtin_ia32_pcmpeqb512_mask ((__v64qi) __A, | |
1719 | (__v64qi) __B, | |
1720 | (__mmask64) -1); | |
1721 | } | |
1722 | ||
eee5d6f5 AI |
1723 | extern __inline __mmask64 |
1724 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1725 | _mm512_mask_cmpeq_epu8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1726 | { | |
1727 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1728 | (__v64qi) __B, 0, | |
1729 | __U); | |
1730 | } | |
1731 | ||
936c0fe4 AI |
1732 | extern __inline __mmask64 |
1733 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1734 | _mm512_mask_cmpeq_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1735 | { | |
1736 | return (__mmask64) __builtin_ia32_pcmpeqb512_mask ((__v64qi) __A, | |
1737 | (__v64qi) __B, | |
1738 | __U); | |
1739 | } | |
1740 | ||
eee5d6f5 AI |
1741 | extern __inline __mmask32 |
1742 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1743 | _mm512_cmpeq_epu16_mask (__m512i __A, __m512i __B) | |
1744 | { | |
1745 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1746 | (__v32hi) __B, 0, | |
1747 | (__mmask32) -1); | |
1748 | } | |
1749 | ||
936c0fe4 AI |
1750 | extern __inline __mmask32 |
1751 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1752 | _mm512_cmpeq_epi16_mask (__m512i __A, __m512i __B) | |
1753 | { | |
1754 | return (__mmask32) __builtin_ia32_pcmpeqw512_mask ((__v32hi) __A, | |
1755 | (__v32hi) __B, | |
1756 | (__mmask32) -1); | |
1757 | } | |
1758 | ||
eee5d6f5 AI |
1759 | extern __inline __mmask32 |
1760 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1761 | _mm512_mask_cmpeq_epu16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1762 | { | |
1763 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1764 | (__v32hi) __B, 0, | |
1765 | __U); | |
1766 | } | |
1767 | ||
936c0fe4 AI |
1768 | extern __inline __mmask32 |
1769 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1770 | _mm512_mask_cmpeq_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1771 | { | |
1772 | return (__mmask32) __builtin_ia32_pcmpeqw512_mask ((__v32hi) __A, | |
1773 | (__v32hi) __B, | |
1774 | __U); | |
1775 | } | |
1776 | ||
eee5d6f5 AI |
1777 | extern __inline __mmask64 |
1778 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1779 | _mm512_cmpgt_epu8_mask (__m512i __A, __m512i __B) | |
1780 | { | |
1781 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1782 | (__v64qi) __B, 6, | |
1783 | (__mmask64) -1); | |
1784 | } | |
1785 | ||
936c0fe4 AI |
1786 | extern __inline __mmask64 |
1787 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1788 | _mm512_cmpgt_epi8_mask (__m512i __A, __m512i __B) | |
1789 | { | |
1790 | return (__mmask64) __builtin_ia32_pcmpgtb512_mask ((__v64qi) __A, | |
1791 | (__v64qi) __B, | |
1792 | (__mmask64) -1); | |
1793 | } | |
1794 | ||
eee5d6f5 AI |
1795 | extern __inline __mmask64 |
1796 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1797 | _mm512_mask_cmpgt_epu8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1798 | { | |
1799 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, | |
1800 | (__v64qi) __B, 6, | |
1801 | __U); | |
1802 | } | |
1803 | ||
936c0fe4 AI |
1804 | extern __inline __mmask64 |
1805 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1806 | _mm512_mask_cmpgt_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1807 | { | |
1808 | return (__mmask64) __builtin_ia32_pcmpgtb512_mask ((__v64qi) __A, | |
1809 | (__v64qi) __B, | |
1810 | __U); | |
1811 | } | |
1812 | ||
eee5d6f5 AI |
1813 | extern __inline __mmask32 |
1814 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1815 | _mm512_cmpgt_epu16_mask (__m512i __A, __m512i __B) | |
1816 | { | |
1817 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1818 | (__v32hi) __B, 6, | |
1819 | (__mmask32) -1); | |
1820 | } | |
1821 | ||
936c0fe4 AI |
1822 | extern __inline __mmask32 |
1823 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1824 | _mm512_cmpgt_epi16_mask (__m512i __A, __m512i __B) | |
1825 | { | |
1826 | return (__mmask32) __builtin_ia32_pcmpgtw512_mask ((__v32hi) __A, | |
1827 | (__v32hi) __B, | |
1828 | (__mmask32) -1); | |
1829 | } | |
1830 | ||
eee5d6f5 AI |
1831 | extern __inline __mmask32 |
1832 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1833 | _mm512_mask_cmpgt_epu16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1834 | { | |
1835 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, | |
1836 | (__v32hi) __B, 6, | |
1837 | __U); | |
1838 | } | |
1839 | ||
936c0fe4 AI |
1840 | extern __inline __mmask32 |
1841 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1842 | _mm512_mask_cmpgt_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1843 | { | |
1844 | return (__mmask32) __builtin_ia32_pcmpgtw512_mask ((__v32hi) __A, | |
1845 | (__v32hi) __B, | |
1846 | __U); | |
1847 | } | |
1848 | ||
1849 | extern __inline __mmask64 | |
1850 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1851 | _mm512_movepi8_mask (__m512i __A) | |
1852 | { | |
1853 | return (__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A); | |
1854 | } | |
1855 | ||
1856 | extern __inline __mmask32 | |
1857 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1858 | _mm512_movepi16_mask (__m512i __A) | |
1859 | { | |
1860 | return (__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A); | |
1861 | } | |
1862 | ||
1863 | extern __inline __m512i | |
1864 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1865 | _mm512_movm_epi8 (__mmask64 __A) | |
1866 | { | |
1867 | return (__m512i) __builtin_ia32_cvtmask2b512 (__A); | |
1868 | } | |
1869 | ||
1870 | extern __inline __m512i | |
1871 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1872 | _mm512_movm_epi16 (__mmask32 __A) | |
1873 | { | |
1874 | return (__m512i) __builtin_ia32_cvtmask2w512 (__A); | |
1875 | } | |
1876 | ||
1877 | extern __inline __mmask64 | |
1878 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1879 | _mm512_test_epi8_mask (__m512i __A, __m512i __B) | |
1880 | { | |
1881 | return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A, | |
1882 | (__v64qi) __B, | |
1883 | (__mmask64) -1); | |
1884 | } | |
1885 | ||
1886 | extern __inline __mmask64 | |
1887 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1888 | _mm512_mask_test_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1889 | { | |
1890 | return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A, | |
1891 | (__v64qi) __B, __U); | |
1892 | } | |
1893 | ||
1894 | extern __inline __mmask32 | |
1895 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1896 | _mm512_test_epi16_mask (__m512i __A, __m512i __B) | |
1897 | { | |
1898 | return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A, | |
1899 | (__v32hi) __B, | |
1900 | (__mmask32) -1); | |
1901 | } | |
1902 | ||
1903 | extern __inline __mmask32 | |
1904 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1905 | _mm512_mask_test_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1906 | { | |
1907 | return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A, | |
1908 | (__v32hi) __B, __U); | |
1909 | } | |
1910 | ||
1911 | extern __inline __mmask64 | |
1912 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1913 | _mm512_testn_epi8_mask (__m512i __A, __m512i __B) | |
1914 | { | |
1915 | return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A, | |
1916 | (__v64qi) __B, | |
1917 | (__mmask64) -1); | |
1918 | } | |
1919 | ||
1920 | extern __inline __mmask64 | |
1921 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1922 | _mm512_mask_testn_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) | |
1923 | { | |
1924 | return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A, | |
1925 | (__v64qi) __B, __U); | |
1926 | } | |
1927 | ||
1928 | extern __inline __mmask32 | |
1929 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1930 | _mm512_testn_epi16_mask (__m512i __A, __m512i __B) | |
1931 | { | |
1932 | return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A, | |
1933 | (__v32hi) __B, | |
1934 | (__mmask32) -1); | |
1935 | } | |
1936 | ||
1937 | extern __inline __mmask32 | |
1938 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1939 | _mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) | |
1940 | { | |
1941 | return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A, | |
1942 | (__v32hi) __B, __U); | |
1943 | } | |
1944 | ||
1945 | extern __inline __m512i | |
1946 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1947 | _mm512_shuffle_epi8 (__m512i __A, __m512i __B) | |
1948 | { | |
1949 | return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, | |
1950 | (__v64qi) __B, | |
1951 | (__v64qi) | |
f0515366 | 1952 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1953 | (__mmask64) -1); |
1954 | } | |
1955 | ||
1956 | extern __inline __m512i | |
1957 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1958 | _mm512_mask_shuffle_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
1959 | __m512i __B) | |
1960 | { | |
1961 | return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, | |
1962 | (__v64qi) __B, | |
1963 | (__v64qi) __W, | |
1964 | (__mmask64) __U); | |
1965 | } | |
1966 | ||
1967 | extern __inline __m512i | |
1968 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1969 | _mm512_maskz_shuffle_epi8 (__mmask64 __U, __m512i __A, __m512i __B) | |
1970 | { | |
1971 | return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, | |
1972 | (__v64qi) __B, | |
1973 | (__v64qi) | |
f0515366 | 1974 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1975 | (__mmask64) __U); |
1976 | } | |
1977 | ||
1978 | extern __inline __m512i | |
1979 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1980 | _mm512_min_epu16 (__m512i __A, __m512i __B) | |
1981 | { | |
1982 | return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, | |
1983 | (__v32hi) __B, | |
1984 | (__v32hi) | |
f0515366 | 1985 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1986 | (__mmask32) -1); |
1987 | } | |
1988 | ||
1989 | extern __inline __m512i | |
1990 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
1991 | _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B) | |
1992 | { | |
1993 | return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, | |
1994 | (__v32hi) __B, | |
1995 | (__v32hi) | |
f0515366 | 1996 | _mm512_setzero_si512 (), |
936c0fe4 AI |
1997 | (__mmask32) __M); |
1998 | } | |
1999 | ||
2000 | extern __inline __m512i | |
2001 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2002 | _mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, | |
2003 | __m512i __B) | |
2004 | { | |
2005 | return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, | |
2006 | (__v32hi) __B, | |
2007 | (__v32hi) __W, | |
2008 | (__mmask32) __M); | |
2009 | } | |
2010 | ||
2011 | extern __inline __m512i | |
2012 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2013 | _mm512_min_epi16 (__m512i __A, __m512i __B) | |
2014 | { | |
2015 | return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, | |
2016 | (__v32hi) __B, | |
2017 | (__v32hi) | |
f0515366 | 2018 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2019 | (__mmask32) -1); |
2020 | } | |
2021 | ||
2022 | extern __inline __m512i | |
2023 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2024 | _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B) | |
2025 | { | |
2026 | return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, | |
2027 | (__v32hi) __B, | |
2028 | (__v32hi) | |
f0515366 | 2029 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2030 | (__mmask32) __M); |
2031 | } | |
2032 | ||
2033 | extern __inline __m512i | |
2034 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2035 | _mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, | |
2036 | __m512i __B) | |
2037 | { | |
2038 | return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, | |
2039 | (__v32hi) __B, | |
2040 | (__v32hi) __W, | |
2041 | (__mmask32) __M); | |
2042 | } | |
2043 | ||
2044 | extern __inline __m512i | |
2045 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2046 | _mm512_max_epu8 (__m512i __A, __m512i __B) | |
2047 | { | |
2048 | return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, | |
2049 | (__v64qi) __B, | |
2050 | (__v64qi) | |
f0515366 | 2051 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2052 | (__mmask64) -1); |
2053 | } | |
2054 | ||
2055 | extern __inline __m512i | |
2056 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2057 | _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B) | |
2058 | { | |
2059 | return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, | |
2060 | (__v64qi) __B, | |
2061 | (__v64qi) | |
f0515366 | 2062 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2063 | (__mmask64) __M); |
2064 | } | |
2065 | ||
2066 | extern __inline __m512i | |
2067 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2068 | _mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, | |
2069 | __m512i __B) | |
2070 | { | |
2071 | return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, | |
2072 | (__v64qi) __B, | |
2073 | (__v64qi) __W, | |
2074 | (__mmask64) __M); | |
2075 | } | |
2076 | ||
2077 | extern __inline __m512i | |
2078 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2079 | _mm512_max_epi8 (__m512i __A, __m512i __B) | |
2080 | { | |
2081 | return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, | |
2082 | (__v64qi) __B, | |
2083 | (__v64qi) | |
f0515366 | 2084 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2085 | (__mmask64) -1); |
2086 | } | |
2087 | ||
2088 | extern __inline __m512i | |
2089 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2090 | _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B) | |
2091 | { | |
2092 | return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, | |
2093 | (__v64qi) __B, | |
2094 | (__v64qi) | |
f0515366 | 2095 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2096 | (__mmask64) __M); |
2097 | } | |
2098 | ||
2099 | extern __inline __m512i | |
2100 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2101 | _mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, | |
2102 | __m512i __B) | |
2103 | { | |
2104 | return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, | |
2105 | (__v64qi) __B, | |
2106 | (__v64qi) __W, | |
2107 | (__mmask64) __M); | |
2108 | } | |
2109 | ||
2110 | extern __inline __m512i | |
2111 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2112 | _mm512_min_epu8 (__m512i __A, __m512i __B) | |
2113 | { | |
2114 | return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, | |
2115 | (__v64qi) __B, | |
2116 | (__v64qi) | |
f0515366 | 2117 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2118 | (__mmask64) -1); |
2119 | } | |
2120 | ||
2121 | extern __inline __m512i | |
2122 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2123 | _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B) | |
2124 | { | |
2125 | return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, | |
2126 | (__v64qi) __B, | |
2127 | (__v64qi) | |
f0515366 | 2128 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2129 | (__mmask64) __M); |
2130 | } | |
2131 | ||
2132 | extern __inline __m512i | |
2133 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2134 | _mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, | |
2135 | __m512i __B) | |
2136 | { | |
2137 | return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, | |
2138 | (__v64qi) __B, | |
2139 | (__v64qi) __W, | |
2140 | (__mmask64) __M); | |
2141 | } | |
2142 | ||
2143 | extern __inline __m512i | |
2144 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2145 | _mm512_min_epi8 (__m512i __A, __m512i __B) | |
2146 | { | |
2147 | return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, | |
2148 | (__v64qi) __B, | |
2149 | (__v64qi) | |
f0515366 | 2150 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2151 | (__mmask64) -1); |
2152 | } | |
2153 | ||
2154 | extern __inline __m512i | |
2155 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2156 | _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B) | |
2157 | { | |
2158 | return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, | |
2159 | (__v64qi) __B, | |
2160 | (__v64qi) | |
f0515366 | 2161 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2162 | (__mmask64) __M); |
2163 | } | |
2164 | ||
2165 | extern __inline __m512i | |
2166 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2167 | _mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, | |
2168 | __m512i __B) | |
2169 | { | |
2170 | return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, | |
2171 | (__v64qi) __B, | |
2172 | (__v64qi) __W, | |
2173 | (__mmask64) __M); | |
2174 | } | |
2175 | ||
2176 | extern __inline __m512i | |
2177 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2178 | _mm512_max_epi16 (__m512i __A, __m512i __B) | |
2179 | { | |
2180 | return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, | |
2181 | (__v32hi) __B, | |
2182 | (__v32hi) | |
f0515366 | 2183 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2184 | (__mmask32) -1); |
2185 | } | |
2186 | ||
2187 | extern __inline __m512i | |
2188 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2189 | _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B) | |
2190 | { | |
2191 | return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, | |
2192 | (__v32hi) __B, | |
2193 | (__v32hi) | |
f0515366 | 2194 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2195 | (__mmask32) __M); |
2196 | } | |
2197 | ||
2198 | extern __inline __m512i | |
2199 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2200 | _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A, | |
2201 | __m512i __B) | |
2202 | { | |
2203 | return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, | |
2204 | (__v32hi) __B, | |
2205 | (__v32hi) __W, | |
2206 | (__mmask32) __M); | |
2207 | } | |
2208 | ||
2209 | extern __inline __m512i | |
2210 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2211 | _mm512_max_epu16 (__m512i __A, __m512i __B) | |
2212 | { | |
2213 | return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, | |
2214 | (__v32hi) __B, | |
2215 | (__v32hi) | |
f0515366 | 2216 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2217 | (__mmask32) -1); |
2218 | } | |
2219 | ||
2220 | extern __inline __m512i | |
2221 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2222 | _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B) | |
2223 | { | |
2224 | return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, | |
2225 | (__v32hi) __B, | |
2226 | (__v32hi) | |
f0515366 | 2227 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2228 | (__mmask32) __M); |
2229 | } | |
2230 | ||
2231 | extern __inline __m512i | |
2232 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2233 | _mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, | |
2234 | __m512i __B) | |
2235 | { | |
2236 | return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, | |
2237 | (__v32hi) __B, | |
2238 | (__v32hi) __W, | |
2239 | (__mmask32) __M); | |
2240 | } | |
2241 | ||
2242 | extern __inline __m512i | |
2243 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2244 | _mm512_sra_epi16 (__m512i __A, __m128i __B) | |
2245 | { | |
2246 | return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, | |
2247 | (__v8hi) __B, | |
2248 | (__v32hi) | |
f0515366 | 2249 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2250 | (__mmask32) -1); |
2251 | } | |
2252 | ||
2253 | extern __inline __m512i | |
2254 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2255 | _mm512_mask_sra_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2256 | __m128i __B) | |
2257 | { | |
2258 | return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, | |
2259 | (__v8hi) __B, | |
2260 | (__v32hi) __W, | |
2261 | (__mmask32) __U); | |
2262 | } | |
2263 | ||
2264 | extern __inline __m512i | |
2265 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2266 | _mm512_maskz_sra_epi16 (__mmask32 __U, __m512i __A, __m128i __B) | |
2267 | { | |
2268 | return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, | |
2269 | (__v8hi) __B, | |
2270 | (__v32hi) | |
f0515366 | 2271 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2272 | (__mmask32) __U); |
2273 | } | |
2274 | ||
2275 | extern __inline __m512i | |
2276 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2277 | _mm512_srav_epi16 (__m512i __A, __m512i __B) | |
2278 | { | |
2279 | return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, | |
2280 | (__v32hi) __B, | |
2281 | (__v32hi) | |
f0515366 | 2282 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2283 | (__mmask32) -1); |
2284 | } | |
2285 | ||
2286 | extern __inline __m512i | |
2287 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2288 | _mm512_mask_srav_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2289 | __m512i __B) | |
2290 | { | |
2291 | return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, | |
2292 | (__v32hi) __B, | |
2293 | (__v32hi) __W, | |
2294 | (__mmask32) __U); | |
2295 | } | |
2296 | ||
2297 | extern __inline __m512i | |
2298 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2299 | _mm512_maskz_srav_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
2300 | { | |
2301 | return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, | |
2302 | (__v32hi) __B, | |
2303 | (__v32hi) | |
f0515366 | 2304 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2305 | (__mmask32) __U); |
2306 | } | |
2307 | ||
2308 | extern __inline __m512i | |
2309 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2310 | _mm512_srlv_epi16 (__m512i __A, __m512i __B) | |
2311 | { | |
2312 | return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, | |
2313 | (__v32hi) __B, | |
2314 | (__v32hi) | |
f0515366 | 2315 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2316 | (__mmask32) -1); |
2317 | } | |
2318 | ||
2319 | extern __inline __m512i | |
2320 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2321 | _mm512_mask_srlv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2322 | __m512i __B) | |
2323 | { | |
2324 | return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, | |
2325 | (__v32hi) __B, | |
2326 | (__v32hi) __W, | |
2327 | (__mmask32) __U); | |
2328 | } | |
2329 | ||
2330 | extern __inline __m512i | |
2331 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2332 | _mm512_maskz_srlv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
2333 | { | |
2334 | return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, | |
2335 | (__v32hi) __B, | |
2336 | (__v32hi) | |
f0515366 | 2337 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2338 | (__mmask32) __U); |
2339 | } | |
2340 | ||
2341 | extern __inline __m512i | |
2342 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2343 | _mm512_sllv_epi16 (__m512i __A, __m512i __B) | |
2344 | { | |
2345 | return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, | |
2346 | (__v32hi) __B, | |
2347 | (__v32hi) | |
f0515366 | 2348 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2349 | (__mmask32) -1); |
2350 | } | |
2351 | ||
2352 | extern __inline __m512i | |
2353 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2354 | _mm512_mask_sllv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
2355 | __m512i __B) | |
2356 | { | |
2357 | return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, | |
2358 | (__v32hi) __B, | |
2359 | (__v32hi) __W, | |
2360 | (__mmask32) __U); | |
2361 | } | |
2362 | ||
2363 | extern __inline __m512i | |
2364 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2365 | _mm512_maskz_sllv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) | |
2366 | { | |
2367 | return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, | |
2368 | (__v32hi) __B, | |
2369 | (__v32hi) | |
f0515366 | 2370 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2371 | (__mmask32) __U); |
2372 | } | |
2373 | ||
2374 | extern __inline __m512i | |
2375 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2376 | _mm512_mask_packs_epi16 (__m512i __W, __mmask64 __M, __m512i __A, | |
2377 | __m512i __B) | |
2378 | { | |
2379 | return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, | |
2380 | (__v32hi) __B, | |
2381 | (__v64qi) __W, | |
2382 | (__mmask64) __M); | |
2383 | } | |
2384 | ||
2385 | extern __inline __m512i | |
2386 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2387 | _mm512_maskz_packs_epi16 (__mmask64 __M, __m512i __A, __m512i __B) | |
2388 | { | |
2389 | return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, | |
2390 | (__v32hi) __B, | |
2391 | (__v64qi) | |
f0515366 | 2392 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2393 | __M); |
2394 | } | |
2395 | ||
2396 | extern __inline __m512i | |
2397 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2398 | _mm512_packus_epi16 (__m512i __A, __m512i __B) | |
2399 | { | |
2400 | return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, | |
2401 | (__v32hi) __B, | |
2402 | (__v64qi) | |
f0515366 | 2403 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2404 | (__mmask64) -1); |
2405 | } | |
2406 | ||
2407 | extern __inline __m512i | |
2408 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2409 | _mm512_mask_packus_epi16 (__m512i __W, __mmask64 __M, __m512i __A, | |
2410 | __m512i __B) | |
2411 | { | |
2412 | return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, | |
2413 | (__v32hi) __B, | |
2414 | (__v64qi) __W, | |
2415 | (__mmask64) __M); | |
2416 | } | |
2417 | ||
2418 | extern __inline __m512i | |
2419 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2420 | _mm512_maskz_packus_epi16 (__mmask64 __M, __m512i __A, __m512i __B) | |
2421 | { | |
2422 | return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, | |
2423 | (__v32hi) __B, | |
2424 | (__v64qi) | |
f0515366 | 2425 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2426 | (__mmask64) __M); |
2427 | } | |
2428 | ||
2429 | extern __inline __m512i | |
2430 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2431 | _mm512_abs_epi8 (__m512i __A) | |
2432 | { | |
2433 | return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, | |
2434 | (__v64qi) | |
f0515366 | 2435 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2436 | (__mmask64) -1); |
2437 | } | |
2438 | ||
2439 | extern __inline __m512i | |
2440 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2441 | _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A) | |
2442 | { | |
2443 | return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, | |
2444 | (__v64qi) __W, | |
2445 | (__mmask64) __U); | |
2446 | } | |
2447 | ||
2448 | extern __inline __m512i | |
2449 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2450 | _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A) | |
2451 | { | |
2452 | return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, | |
2453 | (__v64qi) | |
f0515366 | 2454 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2455 | (__mmask64) __U); |
2456 | } | |
2457 | ||
2458 | extern __inline __m512i | |
2459 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2460 | _mm512_abs_epi16 (__m512i __A) | |
2461 | { | |
2462 | return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, | |
2463 | (__v32hi) | |
f0515366 | 2464 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2465 | (__mmask32) -1); |
2466 | } | |
2467 | ||
2468 | extern __inline __m512i | |
2469 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2470 | _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A) | |
2471 | { | |
2472 | return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, | |
2473 | (__v32hi) __W, | |
2474 | (__mmask32) __U); | |
2475 | } | |
2476 | ||
2477 | extern __inline __m512i | |
2478 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2479 | _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A) | |
2480 | { | |
2481 | return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, | |
2482 | (__v32hi) | |
f0515366 | 2483 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2484 | (__mmask32) __U); |
2485 | } | |
2486 | ||
eee5d6f5 AI |
2487 | extern __inline __mmask64 |
2488 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2489 | _mm512_mask_cmpneq_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2490 | { | |
2491 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2492 | (__v64qi) __Y, 4, | |
2493 | (__mmask64) __M); | |
2494 | } | |
2495 | ||
2496 | extern __inline __mmask64 | |
2497 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2498 | _mm512_mask_cmplt_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2499 | { | |
2500 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2501 | (__v64qi) __Y, 1, | |
2502 | (__mmask64) __M); | |
2503 | } | |
2504 | ||
2505 | extern __inline __mmask64 | |
2506 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2507 | _mm512_mask_cmpge_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2508 | { | |
2509 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2510 | (__v64qi) __Y, 5, | |
2511 | (__mmask64) __M); | |
2512 | } | |
2513 | ||
2514 | extern __inline __mmask64 | |
2515 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2516 | _mm512_mask_cmple_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2517 | { | |
2518 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2519 | (__v64qi) __Y, 2, | |
2520 | (__mmask64) __M); | |
2521 | } | |
2522 | ||
2523 | extern __inline __mmask32 | |
2524 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2525 | _mm512_mask_cmpneq_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2526 | { | |
2527 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2528 | (__v32hi) __Y, 4, | |
2529 | (__mmask32) __M); | |
2530 | } | |
2531 | ||
2532 | extern __inline __mmask32 | |
2533 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2534 | _mm512_mask_cmplt_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2535 | { | |
2536 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2537 | (__v32hi) __Y, 1, | |
2538 | (__mmask32) __M); | |
2539 | } | |
2540 | ||
2541 | extern __inline __mmask32 | |
2542 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2543 | _mm512_mask_cmpge_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2544 | { | |
2545 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2546 | (__v32hi) __Y, 5, | |
2547 | (__mmask32) __M); | |
2548 | } | |
2549 | ||
2550 | extern __inline __mmask32 | |
2551 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2552 | _mm512_mask_cmple_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2553 | { | |
2554 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2555 | (__v32hi) __Y, 2, | |
2556 | (__mmask32) __M); | |
2557 | } | |
2558 | ||
2559 | extern __inline __mmask64 | |
2560 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2561 | _mm512_mask_cmpneq_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2562 | { | |
2563 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2564 | (__v64qi) __Y, 4, | |
2565 | (__mmask64) __M); | |
2566 | } | |
2567 | ||
2568 | extern __inline __mmask64 | |
2569 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2570 | _mm512_mask_cmplt_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2571 | { | |
2572 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2573 | (__v64qi) __Y, 1, | |
2574 | (__mmask64) __M); | |
2575 | } | |
2576 | ||
2577 | extern __inline __mmask64 | |
2578 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2579 | _mm512_mask_cmpge_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2580 | { | |
2581 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2582 | (__v64qi) __Y, 5, | |
2583 | (__mmask64) __M); | |
2584 | } | |
2585 | ||
2586 | extern __inline __mmask64 | |
2587 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2588 | _mm512_mask_cmple_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) | |
2589 | { | |
2590 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2591 | (__v64qi) __Y, 2, | |
2592 | (__mmask64) __M); | |
2593 | } | |
2594 | ||
2595 | extern __inline __mmask32 | |
2596 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2597 | _mm512_mask_cmpneq_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2598 | { | |
2599 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2600 | (__v32hi) __Y, 4, | |
2601 | (__mmask32) __M); | |
2602 | } | |
2603 | ||
2604 | extern __inline __mmask32 | |
2605 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2606 | _mm512_mask_cmplt_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2607 | { | |
2608 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2609 | (__v32hi) __Y, 1, | |
2610 | (__mmask32) __M); | |
2611 | } | |
2612 | ||
2613 | extern __inline __mmask32 | |
2614 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2615 | _mm512_mask_cmpge_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2616 | { | |
2617 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2618 | (__v32hi) __Y, 5, | |
2619 | (__mmask32) __M); | |
2620 | } | |
2621 | ||
2622 | extern __inline __mmask32 | |
2623 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2624 | _mm512_mask_cmple_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) | |
2625 | { | |
2626 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2627 | (__v32hi) __Y, 2, | |
2628 | (__mmask32) __M); | |
2629 | } | |
2630 | ||
2631 | extern __inline __mmask64 | |
2632 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2633 | _mm512_cmpneq_epu8_mask (__m512i __X, __m512i __Y) | |
2634 | { | |
2635 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2636 | (__v64qi) __Y, 4, | |
281f7120 | 2637 | (__mmask64) -1); |
eee5d6f5 AI |
2638 | } |
2639 | ||
2640 | extern __inline __mmask64 | |
2641 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2642 | _mm512_cmplt_epu8_mask (__m512i __X, __m512i __Y) | |
2643 | { | |
2644 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2645 | (__v64qi) __Y, 1, | |
281f7120 | 2646 | (__mmask64) -1); |
eee5d6f5 AI |
2647 | } |
2648 | ||
2649 | extern __inline __mmask64 | |
2650 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2651 | _mm512_cmpge_epu8_mask (__m512i __X, __m512i __Y) | |
2652 | { | |
2653 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2654 | (__v64qi) __Y, 5, | |
281f7120 | 2655 | (__mmask64) -1); |
eee5d6f5 AI |
2656 | } |
2657 | ||
2658 | extern __inline __mmask64 | |
2659 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2660 | _mm512_cmple_epu8_mask (__m512i __X, __m512i __Y) | |
2661 | { | |
2662 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
2663 | (__v64qi) __Y, 2, | |
281f7120 | 2664 | (__mmask64) -1); |
eee5d6f5 AI |
2665 | } |
2666 | ||
2667 | extern __inline __mmask32 | |
2668 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2669 | _mm512_cmpneq_epu16_mask (__m512i __X, __m512i __Y) | |
2670 | { | |
2671 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2672 | (__v32hi) __Y, 4, | |
281f7120 | 2673 | (__mmask32) -1); |
eee5d6f5 AI |
2674 | } |
2675 | ||
2676 | extern __inline __mmask32 | |
2677 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2678 | _mm512_cmplt_epu16_mask (__m512i __X, __m512i __Y) | |
2679 | { | |
2680 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2681 | (__v32hi) __Y, 1, | |
281f7120 | 2682 | (__mmask32) -1); |
eee5d6f5 AI |
2683 | } |
2684 | ||
2685 | extern __inline __mmask32 | |
2686 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2687 | _mm512_cmpge_epu16_mask (__m512i __X, __m512i __Y) | |
2688 | { | |
2689 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2690 | (__v32hi) __Y, 5, | |
281f7120 | 2691 | (__mmask32) -1); |
eee5d6f5 AI |
2692 | } |
2693 | ||
2694 | extern __inline __mmask32 | |
2695 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2696 | _mm512_cmple_epu16_mask (__m512i __X, __m512i __Y) | |
2697 | { | |
2698 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
2699 | (__v32hi) __Y, 2, | |
281f7120 | 2700 | (__mmask32) -1); |
eee5d6f5 AI |
2701 | } |
2702 | ||
2703 | extern __inline __mmask64 | |
2704 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2705 | _mm512_cmpneq_epi8_mask (__m512i __X, __m512i __Y) | |
2706 | { | |
2707 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2708 | (__v64qi) __Y, 4, | |
281f7120 | 2709 | (__mmask64) -1); |
eee5d6f5 AI |
2710 | } |
2711 | ||
2712 | extern __inline __mmask64 | |
2713 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2714 | _mm512_cmplt_epi8_mask (__m512i __X, __m512i __Y) | |
2715 | { | |
2716 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2717 | (__v64qi) __Y, 1, | |
281f7120 | 2718 | (__mmask64) -1); |
eee5d6f5 AI |
2719 | } |
2720 | ||
2721 | extern __inline __mmask64 | |
2722 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2723 | _mm512_cmpge_epi8_mask (__m512i __X, __m512i __Y) | |
2724 | { | |
2725 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2726 | (__v64qi) __Y, 5, | |
281f7120 | 2727 | (__mmask64) -1); |
eee5d6f5 AI |
2728 | } |
2729 | ||
2730 | extern __inline __mmask64 | |
2731 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2732 | _mm512_cmple_epi8_mask (__m512i __X, __m512i __Y) | |
2733 | { | |
2734 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
2735 | (__v64qi) __Y, 2, | |
281f7120 | 2736 | (__mmask64) -1); |
eee5d6f5 AI |
2737 | } |
2738 | ||
2739 | extern __inline __mmask32 | |
2740 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2741 | _mm512_cmpneq_epi16_mask (__m512i __X, __m512i __Y) | |
2742 | { | |
2743 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2744 | (__v32hi) __Y, 4, | |
281f7120 | 2745 | (__mmask32) -1); |
eee5d6f5 AI |
2746 | } |
2747 | ||
2748 | extern __inline __mmask32 | |
2749 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2750 | _mm512_cmplt_epi16_mask (__m512i __X, __m512i __Y) | |
2751 | { | |
2752 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2753 | (__v32hi) __Y, 1, | |
281f7120 | 2754 | (__mmask32) -1); |
eee5d6f5 AI |
2755 | } |
2756 | ||
2757 | extern __inline __mmask32 | |
2758 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2759 | _mm512_cmpge_epi16_mask (__m512i __X, __m512i __Y) | |
2760 | { | |
2761 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2762 | (__v32hi) __Y, 5, | |
281f7120 | 2763 | (__mmask32) -1); |
eee5d6f5 AI |
2764 | } |
2765 | ||
2766 | extern __inline __mmask32 | |
2767 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2768 | _mm512_cmple_epi16_mask (__m512i __X, __m512i __Y) | |
2769 | { | |
2770 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
2771 | (__v32hi) __Y, 2, | |
281f7120 | 2772 | (__mmask32) -1); |
eee5d6f5 AI |
2773 | } |
2774 | ||
030fc94a JJ |
2775 | extern __inline __m512i |
2776 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2777 | _mm512_packs_epi32 (__m512i __A, __m512i __B) | |
2778 | { | |
2779 | return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, | |
2780 | (__v16si) __B, | |
2781 | (__v32hi) | |
2782 | _mm512_setzero_si512 (), | |
2783 | (__mmask32) -1); | |
2784 | } | |
2785 | ||
2786 | extern __inline __m512i | |
2787 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2788 | _mm512_maskz_packs_epi32 (__mmask32 __M, __m512i __A, __m512i __B) | |
2789 | { | |
2790 | return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, | |
2791 | (__v16si) __B, | |
2792 | (__v32hi) | |
2793 | _mm512_setzero_si512 (), | |
2794 | __M); | |
2795 | } | |
2796 | ||
2797 | extern __inline __m512i | |
2798 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2799 | _mm512_mask_packs_epi32 (__m512i __W, __mmask32 __M, __m512i __A, | |
2800 | __m512i __B) | |
2801 | { | |
2802 | return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, | |
2803 | (__v16si) __B, | |
2804 | (__v32hi) __W, | |
2805 | __M); | |
2806 | } | |
2807 | ||
2808 | extern __inline __m512i | |
2809 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2810 | _mm512_packus_epi32 (__m512i __A, __m512i __B) | |
2811 | { | |
2812 | return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, | |
2813 | (__v16si) __B, | |
2814 | (__v32hi) | |
2815 | _mm512_setzero_si512 (), | |
2816 | (__mmask32) -1); | |
2817 | } | |
2818 | ||
2819 | extern __inline __m512i | |
2820 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2821 | _mm512_maskz_packus_epi32 (__mmask32 __M, __m512i __A, __m512i __B) | |
2822 | { | |
2823 | return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, | |
2824 | (__v16si) __B, | |
2825 | (__v32hi) | |
2826 | _mm512_setzero_si512 (), | |
2827 | __M); | |
2828 | } | |
2829 | ||
2830 | extern __inline __m512i | |
2831 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2832 | _mm512_mask_packus_epi32 (__m512i __W, __mmask32 __M, __m512i __A, | |
2833 | __m512i __B) | |
2834 | { | |
2835 | return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, | |
2836 | (__v16si) __B, | |
2837 | (__v32hi) __W, | |
2838 | __M); | |
2839 | } | |
2840 | ||
936c0fe4 | 2841 | #ifdef __OPTIMIZE__ |
d8ea3e7c AS |
2842 | extern __inline __mmask64 |
2843 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2844 | _kshiftli_mask64 (__mmask64 __A, unsigned int __B) | |
2845 | { | |
2846 | return (__mmask64) __builtin_ia32_kshiftlidi ((__mmask64) __A, | |
2847 | (__mmask8) __B); | |
2848 | } | |
2849 | ||
d8ea3e7c AS |
2850 | extern __inline __mmask64 |
2851 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2852 | _kshiftri_mask64 (__mmask64 __A, unsigned int __B) | |
2853 | { | |
2854 | return (__mmask64) __builtin_ia32_kshiftridi ((__mmask64) __A, | |
2855 | (__mmask8) __B); | |
2856 | } | |
2857 | ||
936c0fe4 AI |
2858 | extern __inline __m512i |
2859 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2860 | _mm512_alignr_epi8 (__m512i __A, __m512i __B, const int __N) | |
2861 | { | |
2862 | return (__m512i) __builtin_ia32_palignr512 ((__v8di) __A, | |
2863 | (__v8di) __B, __N * 8); | |
2864 | } | |
2865 | ||
2866 | extern __inline __m512i | |
2867 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2868 | _mm512_mask_alignr_epi8 (__m512i __W, __mmask64 __U, __m512i __A, | |
2869 | __m512i __B, const int __N) | |
2870 | { | |
2871 | return (__m512i) __builtin_ia32_palignr512_mask ((__v8di) __A, | |
2872 | (__v8di) __B, | |
2873 | __N * 8, | |
2874 | (__v8di) __W, | |
2875 | (__mmask64) __U); | |
2876 | } | |
2877 | ||
2878 | extern __inline __m512i | |
2879 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2880 | _mm512_maskz_alignr_epi8 (__mmask64 __U, __m512i __A, __m512i __B, | |
2881 | const int __N) | |
2882 | { | |
2883 | return (__m512i) __builtin_ia32_palignr512_mask ((__v8di) __A, | |
2884 | (__v8di) __B, | |
2885 | __N * 8, | |
2886 | (__v8di) | |
2887 | _mm512_setzero_si512 (), | |
2888 | (__mmask64) __U); | |
2889 | } | |
2890 | ||
2891 | extern __inline __m512i | |
2892 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2893 | _mm512_dbsad_epu8 (__m512i __A, __m512i __B, const int __imm) | |
2894 | { | |
2895 | return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, | |
2896 | (__v64qi) __B, | |
2897 | __imm, | |
2898 | (__v32hi) | |
f0515366 | 2899 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2900 | (__mmask32) -1); |
2901 | } | |
2902 | ||
2903 | extern __inline __m512i | |
2904 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2905 | _mm512_mask_dbsad_epu8 (__m512i __W, __mmask32 __U, __m512i __A, | |
2906 | __m512i __B, const int __imm) | |
2907 | { | |
2908 | return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, | |
2909 | (__v64qi) __B, | |
2910 | __imm, | |
2911 | (__v32hi) __W, | |
2912 | (__mmask32) __U); | |
2913 | } | |
2914 | ||
2915 | extern __inline __m512i | |
2916 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2917 | _mm512_maskz_dbsad_epu8 (__mmask32 __U, __m512i __A, __m512i __B, | |
2918 | const int __imm) | |
2919 | { | |
2920 | return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, | |
2921 | (__v64qi) __B, | |
2922 | __imm, | |
2923 | (__v32hi) | |
f0515366 | 2924 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2925 | (__mmask32) __U); |
2926 | } | |
2927 | ||
2928 | extern __inline __m512i | |
2929 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4a84a2db | 2930 | _mm512_srli_epi16 (__m512i __A, const unsigned int __imm) |
936c0fe4 AI |
2931 | { |
2932 | return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, | |
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_mask_srli_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
4a84a2db | 2941 | const unsigned int __imm) |
936c0fe4 AI |
2942 | { |
2943 | return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, | |
2944 | (__v32hi) __W, | |
2945 | (__mmask32) __U); | |
2946 | } | |
2947 | ||
2948 | extern __inline __m512i | |
2949 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2950 | _mm512_maskz_srli_epi16 (__mmask32 __U, __m512i __A, const int __imm) | |
2951 | { | |
2952 | return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, | |
2953 | (__v32hi) | |
f0515366 | 2954 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2955 | (__mmask32) __U); |
2956 | } | |
2957 | ||
2958 | extern __inline __m512i | |
2959 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4a84a2db | 2960 | _mm512_slli_epi16 (__m512i __A, const unsigned int __B) |
936c0fe4 AI |
2961 | { |
2962 | return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, | |
2963 | (__v32hi) | |
f0515366 | 2964 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2965 | (__mmask32) -1); |
2966 | } | |
2967 | ||
2968 | extern __inline __m512i | |
2969 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2970 | _mm512_mask_slli_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
4a84a2db | 2971 | const unsigned int __B) |
936c0fe4 AI |
2972 | { |
2973 | return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, | |
2974 | (__v32hi) __W, | |
2975 | (__mmask32) __U); | |
2976 | } | |
2977 | ||
2978 | extern __inline __m512i | |
2979 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4a84a2db | 2980 | _mm512_maskz_slli_epi16 (__mmask32 __U, __m512i __A, const unsigned int __B) |
936c0fe4 AI |
2981 | { |
2982 | return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, | |
2983 | (__v32hi) | |
f0515366 | 2984 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2985 | (__mmask32) __U); |
2986 | } | |
2987 | ||
2988 | extern __inline __m512i | |
2989 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
2990 | _mm512_shufflehi_epi16 (__m512i __A, const int __imm) | |
2991 | { | |
2992 | return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, | |
2993 | __imm, | |
2994 | (__v32hi) | |
f0515366 | 2995 | _mm512_setzero_si512 (), |
936c0fe4 AI |
2996 | (__mmask32) -1); |
2997 | } | |
2998 | ||
2999 | extern __inline __m512i | |
3000 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3001 | _mm512_mask_shufflehi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
3002 | const int __imm) | |
3003 | { | |
3004 | return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, | |
3005 | __imm, | |
3006 | (__v32hi) __W, | |
3007 | (__mmask32) __U); | |
3008 | } | |
3009 | ||
3010 | extern __inline __m512i | |
3011 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3012 | _mm512_maskz_shufflehi_epi16 (__mmask32 __U, __m512i __A, | |
3013 | const int __imm) | |
3014 | { | |
3015 | return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, | |
3016 | __imm, | |
3017 | (__v32hi) | |
f0515366 | 3018 | _mm512_setzero_si512 (), |
936c0fe4 AI |
3019 | (__mmask32) __U); |
3020 | } | |
3021 | ||
3022 | extern __inline __m512i | |
3023 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3024 | _mm512_shufflelo_epi16 (__m512i __A, const int __imm) | |
3025 | { | |
3026 | return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, | |
3027 | __imm, | |
3028 | (__v32hi) | |
f0515366 | 3029 | _mm512_setzero_si512 (), |
936c0fe4 AI |
3030 | (__mmask32) -1); |
3031 | } | |
3032 | ||
3033 | extern __inline __m512i | |
3034 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3035 | _mm512_mask_shufflelo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
3036 | const int __imm) | |
3037 | { | |
3038 | return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, | |
3039 | __imm, | |
3040 | (__v32hi) __W, | |
3041 | (__mmask32) __U); | |
3042 | } | |
3043 | ||
3044 | extern __inline __m512i | |
3045 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3046 | _mm512_maskz_shufflelo_epi16 (__mmask32 __U, __m512i __A, | |
3047 | const int __imm) | |
3048 | { | |
3049 | return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, | |
3050 | __imm, | |
3051 | (__v32hi) | |
f0515366 | 3052 | _mm512_setzero_si512 (), |
936c0fe4 AI |
3053 | (__mmask32) __U); |
3054 | } | |
3055 | ||
3056 | extern __inline __m512i | |
3057 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4a84a2db | 3058 | _mm512_srai_epi16 (__m512i __A, const unsigned int __imm) |
936c0fe4 AI |
3059 | { |
3060 | return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, | |
3061 | (__v32hi) | |
f0515366 | 3062 | _mm512_setzero_si512 (), |
936c0fe4 AI |
3063 | (__mmask32) -1); |
3064 | } | |
3065 | ||
3066 | extern __inline __m512i | |
3067 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3068 | _mm512_mask_srai_epi16 (__m512i __W, __mmask32 __U, __m512i __A, | |
4a84a2db | 3069 | const unsigned int __imm) |
936c0fe4 AI |
3070 | { |
3071 | return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, | |
3072 | (__v32hi) __W, | |
3073 | (__mmask32) __U); | |
3074 | } | |
3075 | ||
3076 | extern __inline __m512i | |
3077 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
4a84a2db | 3078 | _mm512_maskz_srai_epi16 (__mmask32 __U, __m512i __A, const unsigned int __imm) |
936c0fe4 AI |
3079 | { |
3080 | return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, | |
3081 | (__v32hi) | |
f0515366 | 3082 | _mm512_setzero_si512 (), |
936c0fe4 AI |
3083 | (__mmask32) __U); |
3084 | } | |
3085 | ||
3086 | extern __inline __m512i | |
3087 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3088 | _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W) | |
3089 | { | |
3090 | return (__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) __A, | |
3091 | (__v32hi) __W, | |
3092 | (__mmask32) __U); | |
3093 | } | |
3094 | ||
3095 | extern __inline __m512i | |
3096 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3097 | _mm512_mask_blend_epi8 (__mmask64 __U, __m512i __A, __m512i __W) | |
3098 | { | |
3099 | return (__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) __A, | |
3100 | (__v64qi) __W, | |
3101 | (__mmask64) __U); | |
3102 | } | |
3103 | ||
3104 | extern __inline __mmask32 | |
3105 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3106 | _mm512_mask_cmp_epi16_mask (__mmask32 __U, __m512i __X, __m512i __Y, | |
3107 | const int __P) | |
3108 | { | |
3109 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
3110 | (__v32hi) __Y, __P, | |
3111 | (__mmask32) __U); | |
3112 | } | |
3113 | ||
3114 | extern __inline __mmask32 | |
3115 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3116 | _mm512_cmp_epi16_mask (__m512i __X, __m512i __Y, const int __P) | |
3117 | { | |
3118 | return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, | |
3119 | (__v32hi) __Y, __P, | |
3120 | (__mmask32) -1); | |
3121 | } | |
3122 | ||
3123 | extern __inline __mmask64 | |
3124 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
18379eea | 3125 | _mm512_mask_cmp_epi8_mask (__mmask64 __U, __m512i __X, __m512i __Y, |
936c0fe4 AI |
3126 | const int __P) |
3127 | { | |
3128 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
3129 | (__v64qi) __Y, __P, | |
3130 | (__mmask64) __U); | |
3131 | } | |
3132 | ||
3133 | extern __inline __mmask64 | |
3134 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3135 | _mm512_cmp_epi8_mask (__m512i __X, __m512i __Y, const int __P) | |
3136 | { | |
3137 | return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, | |
3138 | (__v64qi) __Y, __P, | |
3139 | (__mmask64) -1); | |
3140 | } | |
3141 | ||
3142 | extern __inline __mmask32 | |
3143 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3144 | _mm512_mask_cmp_epu16_mask (__mmask32 __U, __m512i __X, __m512i __Y, | |
3145 | const int __P) | |
3146 | { | |
3147 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
3148 | (__v32hi) __Y, __P, | |
3149 | (__mmask32) __U); | |
3150 | } | |
3151 | ||
3152 | extern __inline __mmask32 | |
3153 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3154 | _mm512_cmp_epu16_mask (__m512i __X, __m512i __Y, const int __P) | |
3155 | { | |
3156 | return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, | |
3157 | (__v32hi) __Y, __P, | |
3158 | (__mmask32) -1); | |
3159 | } | |
3160 | ||
3161 | extern __inline __mmask64 | |
3162 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
18379eea | 3163 | _mm512_mask_cmp_epu8_mask (__mmask64 __U, __m512i __X, __m512i __Y, |
936c0fe4 AI |
3164 | const int __P) |
3165 | { | |
3166 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
3167 | (__v64qi) __Y, __P, | |
3168 | (__mmask64) __U); | |
3169 | } | |
3170 | ||
3171 | extern __inline __mmask64 | |
3172 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3173 | _mm512_cmp_epu8_mask (__m512i __X, __m512i __Y, const int __P) | |
3174 | { | |
3175 | return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, | |
3176 | (__v64qi) __Y, __P, | |
3177 | (__mmask64) -1); | |
3178 | } | |
3179 | ||
936c0fe4 AI |
3180 | extern __inline __m512i |
3181 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3182 | _mm512_bslli_epi128 (__m512i __A, const int __N) | |
3183 | { | |
3184 | return (__m512i) __builtin_ia32_pslldq512 (__A, __N * 8); | |
3185 | } | |
3186 | ||
3187 | extern __inline __m512i | |
3188 | __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) | |
3189 | _mm512_bsrli_epi128 (__m512i __A, const int __N) | |
3190 | { | |
3191 | return (__m512i) __builtin_ia32_psrldq512 (__A, __N * 8); | |
3192 | } | |
3193 | ||
3194 | #else | |
d8ea3e7c AS |
3195 | #define _kshiftli_mask64(X, Y) \ |
3196 | ((__mmask64) __builtin_ia32_kshiftlidi ((__mmask64)(X), (__mmask8)(Y))) | |
3197 | ||
d8ea3e7c AS |
3198 | #define _kshiftri_mask64(X, Y) \ |
3199 | ((__mmask64) __builtin_ia32_kshiftridi ((__mmask64)(X), (__mmask8)(Y))) | |
3200 | ||
936c0fe4 AI |
3201 | #define _mm512_alignr_epi8(X, Y, N) \ |
3202 | ((__m512i) __builtin_ia32_palignr512 ((__v8di)(__m512i)(X), \ | |
3203 | (__v8di)(__m512i)(Y), \ | |
0c8217b1 | 3204 | (int)((N) * 8))) |
936c0fe4 AI |
3205 | |
3206 | #define _mm512_mask_alignr_epi8(W, U, X, Y, N) \ | |
3207 | ((__m512i) __builtin_ia32_palignr512_mask ((__v8di)(__m512i)(X), \ | |
0c8217b1 | 3208 | (__v8di)(__m512i)(Y), (int)((N) * 8), \ |
936c0fe4 AI |
3209 | (__v8di)(__m512i)(W), (__mmask64)(U))) |
3210 | ||
3211 | #define _mm512_maskz_alignr_epi8(U, X, Y, N) \ | |
3212 | ((__m512i) __builtin_ia32_palignr512_mask ((__v8di)(__m512i)(X), \ | |
0c8217b1 | 3213 | (__v8di)(__m512i)(Y), (int)((N) * 8), \ |
f0515366 JJ |
3214 | (__v8di)(__m512i) \ |
3215 | _mm512_setzero_si512 (), \ | |
3216 | (__mmask64)(U))) | |
936c0fe4 AI |
3217 | |
3218 | #define _mm512_dbsad_epu8(X, Y, C) \ | |
3219 | ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ | |
3220 | (__v64qi)(__m512i) (Y), (int) (C), \ | |
f0515366 JJ |
3221 | (__v32hi)(__m512i) \ |
3222 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3223 | (__mmask32)-1)) |
3224 | ||
3225 | #define _mm512_mask_dbsad_epu8(W, U, X, Y, C) \ | |
3226 | ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ | |
3227 | (__v64qi)(__m512i) (Y), (int) (C), \ | |
3228 | (__v32hi)(__m512i)(W), \ | |
3229 | (__mmask32)(U))) | |
3230 | ||
3231 | #define _mm512_maskz_dbsad_epu8(U, X, Y, C) \ | |
3232 | ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ | |
3233 | (__v64qi)(__m512i) (Y), (int) (C), \ | |
f0515366 JJ |
3234 | (__v32hi)(__m512i) \ |
3235 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3236 | (__mmask32)(U))) |
3237 | ||
3238 | #define _mm512_srli_epi16(A, B) \ | |
3239 | ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ | |
4a84a2db | 3240 | (unsigned int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)-1)) |
936c0fe4 AI |
3241 | |
3242 | #define _mm512_mask_srli_epi16(W, U, A, B) \ | |
3243 | ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ | |
4a84a2db | 3244 | (unsigned int)(B), (__v32hi)(__m512i)(W), (__mmask32)(U))) |
936c0fe4 AI |
3245 | |
3246 | #define _mm512_maskz_srli_epi16(U, A, B) \ | |
3247 | ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ | |
f0515366 | 3248 | (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)(U))) |
936c0fe4 | 3249 | |
4a84a2db HL |
3250 | #define _mm512_slli_epi16(X, C) \ |
3251 | ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), \ | |
3252 | (unsigned int)(C), \ | |
3253 | (__v32hi)(__m512i)_mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3254 | (__mmask32)-1)) |
3255 | ||
4a84a2db HL |
3256 | #define _mm512_mask_slli_epi16(W, U, X, C) \ |
3257 | ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), \ | |
3258 | (unsigned int)(C), \ | |
3259 | (__v32hi)(__m512i)(W), \ | |
936c0fe4 AI |
3260 | (__mmask32)(U))) |
3261 | ||
4a84a2db HL |
3262 | #define _mm512_maskz_slli_epi16(U, X, C) \ |
3263 | ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), \ | |
3264 | (unsigned int)(C), \ | |
3265 | (__v32hi)(__m512i)_mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3266 | (__mmask32)(U))) |
3267 | ||
3268 | #define _mm512_shufflehi_epi16(A, B) \ | |
3269 | ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3270 | (__v32hi)(__m512i) \ |
3271 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3272 | (__mmask32)-1)) |
3273 | ||
3274 | #define _mm512_mask_shufflehi_epi16(W, U, A, B) \ | |
3275 | ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
3276 | (__v32hi)(__m512i)(W), \ | |
3277 | (__mmask32)(U))) | |
3278 | ||
3279 | #define _mm512_maskz_shufflehi_epi16(U, A, B) \ | |
3280 | ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3281 | (__v32hi)(__m512i) \ |
3282 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3283 | (__mmask32)(U))) |
3284 | ||
3285 | #define _mm512_shufflelo_epi16(A, B) \ | |
3286 | ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3287 | (__v32hi)(__m512i) \ |
3288 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3289 | (__mmask32)-1)) |
3290 | ||
3291 | #define _mm512_mask_shufflelo_epi16(W, U, A, B) \ | |
3292 | ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
3293 | (__v32hi)(__m512i)(W), \ | |
3294 | (__mmask32)(U))) | |
3295 | ||
3296 | #define _mm512_maskz_shufflelo_epi16(U, A, B) \ | |
3297 | ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ | |
f0515366 JJ |
3298 | (__v32hi)(__m512i) \ |
3299 | _mm512_setzero_si512 (), \ | |
936c0fe4 AI |
3300 | (__mmask32)(U))) |
3301 | ||
3302 | #define _mm512_srai_epi16(A, B) \ | |
3303 | ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ | |
4a84a2db | 3304 | (unsigned int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)-1)) |
936c0fe4 AI |
3305 | |
3306 | #define _mm512_mask_srai_epi16(W, U, A, B) \ | |
3307 | ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ | |
4a84a2db | 3308 | (unsigned int)(B), (__v32hi)(__m512i)(W), (__mmask32)(U))) |
936c0fe4 AI |
3309 | |
3310 | #define _mm512_maskz_srai_epi16(U, A, B) \ | |
3311 | ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ | |
4a84a2db | 3312 | (unsigned int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)(U))) |
936c0fe4 AI |
3313 | |
3314 | #define _mm512_mask_blend_epi16(__U, __A, __W) \ | |
3315 | ((__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) (__A), \ | |
3316 | (__v32hi) (__W), \ | |
3317 | (__mmask32) (__U))) | |
3318 | ||
3319 | #define _mm512_mask_blend_epi8(__U, __A, __W) \ | |
3320 | ((__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) (__A), \ | |
3321 | (__v64qi) (__W), \ | |
3322 | (__mmask64) (__U))) | |
3323 | ||
3324 | #define _mm512_cmp_epi16_mask(X, Y, P) \ | |
3325 | ((__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3326 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3327 | (__mmask32)(-1))) | |
3328 | ||
3329 | #define _mm512_cmp_epi8_mask(X, Y, P) \ | |
3330 | ((__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3331 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3332 | (__mmask64)(-1))) | |
3333 | ||
3334 | #define _mm512_cmp_epu16_mask(X, Y, P) \ | |
3335 | ((__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3336 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3337 | (__mmask32)(-1))) | |
3338 | ||
3339 | #define _mm512_cmp_epu8_mask(X, Y, P) \ | |
3340 | ((__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3341 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3342 | (__mmask64)(-1))) | |
3343 | ||
3344 | #define _mm512_mask_cmp_epi16_mask(M, X, Y, P) \ | |
3345 | ((__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3346 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3347 | (__mmask32)(M))) | |
3348 | ||
3349 | #define _mm512_mask_cmp_epi8_mask(M, X, Y, P) \ | |
3350 | ((__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3351 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3352 | (__mmask64)(M))) | |
3353 | ||
3354 | #define _mm512_mask_cmp_epu16_mask(M, X, Y, P) \ | |
3355 | ((__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi)(__m512i)(X), \ | |
3356 | (__v32hi)(__m512i)(Y), (int)(P),\ | |
3357 | (__mmask32)(M))) | |
3358 | ||
3359 | #define _mm512_mask_cmp_epu8_mask(M, X, Y, P) \ | |
3360 | ((__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi)(__m512i)(X), \ | |
3361 | (__v64qi)(__m512i)(Y), (int)(P),\ | |
3362 | (__mmask64)(M))) | |
3363 | ||
3364 | #define _mm512_bslli_epi128(A, N) \ | |
3365 | ((__m512i)__builtin_ia32_pslldq512 ((__m512i)(A), (int)(N) * 8)) | |
3366 | ||
3367 | #define _mm512_bsrli_epi128(A, N) \ | |
3368 | ((__m512i)__builtin_ia32_psrldq512 ((__m512i)(A), (int)(N) * 8)) | |
3369 | ||
3370 | #endif | |
3371 | ||
03a85048 HJ |
3372 | #ifdef __DISABLE_AVX512BW_512__ |
3373 | #undef __DISABLE_AVX512BW_512__ | |
936c0fe4 | 3374 | #pragma GCC pop_options |
03a85048 | 3375 | #endif /* __DISABLE_AVX512BW_512__ */ |
936c0fe4 AI |
3376 | |
3377 | #endif /* _AVX512BWINTRIN_H_INCLUDED */ |