1 /*===------------- avx512bwintrin.h - AVX512BW intrinsics ------------------===
4 * Permission is hereby granted, free of charge, to any person obtaining a copy
5 * of this software and associated documentation files (the "Software"), to deal
6 * in the Software without restriction, including without limitation the rights
7 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
8 * copies of the Software, and to permit persons to whom the Software is
9 * furnished to do so, subject to the following conditions:
11 * The above copyright notice and this permission notice shall be included in
12 * all copies or substantial portions of the Software.
14 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
15 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
16 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
17 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
18 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
19 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
22 *===-----------------------------------------------------------------------===
25 #error "Never use <avx512bwintrin.h> directly; include <immintrin.h> instead."
28 #ifndef __AVX512BWINTRIN_H
29 #define __AVX512BWINTRIN_H
31 typedef unsigned int __mmask32;
32 typedef unsigned long long __mmask64;
33 typedef char __v64qi __attribute__ ((__vector_size__ (64)));
34 typedef short __v32hi __attribute__ ((__vector_size__ (64)));
36 /* Define the default attributes for the functions in this file. */
37 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__))
39 static __inline __v64qi __DEFAULT_FN_ATTRS
40 _mm512_setzero_qi (void) {
41 return (__v64qi){ 0, 0, 0, 0, 0, 0, 0, 0,
42 0, 0, 0, 0, 0, 0, 0, 0,
43 0, 0, 0, 0, 0, 0, 0, 0,
44 0, 0, 0, 0, 0, 0, 0, 0,
45 0, 0, 0, 0, 0, 0, 0, 0,
46 0, 0, 0, 0, 0, 0, 0, 0,
47 0, 0, 0, 0, 0, 0, 0, 0,
48 0, 0, 0, 0, 0, 0, 0, 0 };
51 static __inline __v32hi __DEFAULT_FN_ATTRS
52 _mm512_setzero_hi (void) {
53 return (__v32hi){ 0, 0, 0, 0, 0, 0, 0, 0,
54 0, 0, 0, 0, 0, 0, 0, 0,
55 0, 0, 0, 0, 0, 0, 0, 0,
56 0, 0, 0, 0, 0, 0, 0, 0 };
61 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
62 _mm512_cmpeq_epi8_mask(__m512i __a, __m512i __b) {
63 return (__mmask64)__builtin_ia32_pcmpeqb512_mask((__v64qi)__a, (__v64qi)__b,
67 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
68 _mm512_mask_cmpeq_epi8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
69 return (__mmask64)__builtin_ia32_pcmpeqb512_mask((__v64qi)__a, (__v64qi)__b,
73 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
74 _mm512_cmpeq_epu8_mask(__m512i __a, __m512i __b) {
75 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 0,
79 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
80 _mm512_mask_cmpeq_epu8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
81 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 0,
85 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
86 _mm512_cmpeq_epi16_mask(__m512i __a, __m512i __b) {
87 return (__mmask32)__builtin_ia32_pcmpeqw512_mask((__v32hi)__a, (__v32hi)__b,
91 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
92 _mm512_mask_cmpeq_epi16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
93 return (__mmask32)__builtin_ia32_pcmpeqw512_mask((__v32hi)__a, (__v32hi)__b,
97 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
98 _mm512_cmpeq_epu16_mask(__m512i __a, __m512i __b) {
99 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 0,
103 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
104 _mm512_mask_cmpeq_epu16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
105 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 0,
109 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
110 _mm512_cmpge_epi8_mask(__m512i __a, __m512i __b) {
111 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 5,
115 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
116 _mm512_mask_cmpge_epi8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
117 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 5,
121 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
122 _mm512_cmpge_epu8_mask(__m512i __a, __m512i __b) {
123 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 5,
127 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
128 _mm512_mask_cmpge_epu8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
129 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 5,
133 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
134 _mm512_cmpge_epi16_mask(__m512i __a, __m512i __b) {
135 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 5,
139 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
140 _mm512_mask_cmpge_epi16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
141 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 5,
145 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
146 _mm512_cmpge_epu16_mask(__m512i __a, __m512i __b) {
147 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 5,
151 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
152 _mm512_mask_cmpge_epu16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
153 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 5,
157 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
158 _mm512_cmpgt_epi8_mask(__m512i __a, __m512i __b) {
159 return (__mmask64)__builtin_ia32_pcmpgtb512_mask((__v64qi)__a, (__v64qi)__b,
163 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
164 _mm512_mask_cmpgt_epi8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
165 return (__mmask64)__builtin_ia32_pcmpgtb512_mask((__v64qi)__a, (__v64qi)__b,
169 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
170 _mm512_cmpgt_epu8_mask(__m512i __a, __m512i __b) {
171 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 6,
175 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
176 _mm512_mask_cmpgt_epu8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
177 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 6,
181 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
182 _mm512_cmpgt_epi16_mask(__m512i __a, __m512i __b) {
183 return (__mmask32)__builtin_ia32_pcmpgtw512_mask((__v32hi)__a, (__v32hi)__b,
187 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
188 _mm512_mask_cmpgt_epi16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
189 return (__mmask32)__builtin_ia32_pcmpgtw512_mask((__v32hi)__a, (__v32hi)__b,
193 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
194 _mm512_cmpgt_epu16_mask(__m512i __a, __m512i __b) {
195 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 6,
199 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
200 _mm512_mask_cmpgt_epu16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
201 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 6,
205 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
206 _mm512_cmple_epi8_mask(__m512i __a, __m512i __b) {
207 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 2,
211 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
212 _mm512_mask_cmple_epi8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
213 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 2,
217 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
218 _mm512_cmple_epu8_mask(__m512i __a, __m512i __b) {
219 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 2,
223 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
224 _mm512_mask_cmple_epu8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
225 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 2,
229 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
230 _mm512_cmple_epi16_mask(__m512i __a, __m512i __b) {
231 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 2,
235 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
236 _mm512_mask_cmple_epi16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
237 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 2,
241 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
242 _mm512_cmple_epu16_mask(__m512i __a, __m512i __b) {
243 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 2,
247 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
248 _mm512_mask_cmple_epu16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
249 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 2,
253 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
254 _mm512_cmplt_epi8_mask(__m512i __a, __m512i __b) {
255 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 1,
259 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
260 _mm512_mask_cmplt_epi8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
261 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 1,
265 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
266 _mm512_cmplt_epu8_mask(__m512i __a, __m512i __b) {
267 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 1,
271 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
272 _mm512_mask_cmplt_epu8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
273 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 1,
277 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
278 _mm512_cmplt_epi16_mask(__m512i __a, __m512i __b) {
279 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 1,
283 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
284 _mm512_mask_cmplt_epi16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
285 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 1,
289 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
290 _mm512_cmplt_epu16_mask(__m512i __a, __m512i __b) {
291 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 1,
295 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
296 _mm512_mask_cmplt_epu16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
297 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 1,
301 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
302 _mm512_cmpneq_epi8_mask(__m512i __a, __m512i __b) {
303 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 4,
307 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
308 _mm512_mask_cmpneq_epi8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
309 return (__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)__a, (__v64qi)__b, 4,
313 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
314 _mm512_cmpneq_epu8_mask(__m512i __a, __m512i __b) {
315 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 4,
319 static __inline__ __mmask64 __DEFAULT_FN_ATTRS
320 _mm512_mask_cmpneq_epu8_mask(__mmask64 __u, __m512i __a, __m512i __b) {
321 return (__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)__a, (__v64qi)__b, 4,
325 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
326 _mm512_cmpneq_epi16_mask(__m512i __a, __m512i __b) {
327 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 4,
331 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
332 _mm512_mask_cmpneq_epi16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
333 return (__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)__a, (__v32hi)__b, 4,
337 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
338 _mm512_cmpneq_epu16_mask(__m512i __a, __m512i __b) {
339 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 4,
343 static __inline__ __mmask32 __DEFAULT_FN_ATTRS
344 _mm512_mask_cmpneq_epu16_mask(__mmask32 __u, __m512i __a, __m512i __b) {
345 return (__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)__a, (__v32hi)__b, 4,
349 static __inline__ __m512i __DEFAULT_FN_ATTRS
350 _mm512_add_epi8 (__m512i __A, __m512i __B) {
351 return (__m512i) ((__v64qi) __A + (__v64qi) __B);
354 static __inline__ __m512i __DEFAULT_FN_ATTRS
355 _mm512_mask_add_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) {
356 return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A,
362 static __inline__ __m512i __DEFAULT_FN_ATTRS
363 _mm512_maskz_add_epi8 (__mmask64 __U, __m512i __A, __m512i __B) {
364 return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A,
367 _mm512_setzero_qi (),
371 static __inline__ __m512i __DEFAULT_FN_ATTRS
372 _mm512_sub_epi8 (__m512i __A, __m512i __B) {
373 return (__m512i) ((__v64qi) __A - (__v64qi) __B);
376 static __inline__ __m512i __DEFAULT_FN_ATTRS
377 _mm512_mask_sub_epi8 (__m512i __W, __mmask64 __U, __m512i __A, __m512i __B) {
378 return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A,
384 static __inline__ __m512i __DEFAULT_FN_ATTRS
385 _mm512_maskz_sub_epi8 (__mmask64 __U, __m512i __A, __m512i __B) {
386 return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A,
389 _mm512_setzero_qi (),
393 static __inline__ __m512i __DEFAULT_FN_ATTRS
394 _mm512_add_epi16 (__m512i __A, __m512i __B) {
395 return (__m512i) ((__v32hi) __A + (__v32hi) __B);
398 static __inline__ __m512i __DEFAULT_FN_ATTRS
399 _mm512_mask_add_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) {
400 return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A,
406 static __inline__ __m512i __DEFAULT_FN_ATTRS
407 _mm512_maskz_add_epi16 (__mmask32 __U, __m512i __A, __m512i __B) {
408 return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A,
411 _mm512_setzero_hi (),
415 static __inline__ __m512i __DEFAULT_FN_ATTRS
416 _mm512_sub_epi16 (__m512i __A, __m512i __B) {
417 return (__m512i) ((__v32hi) __A - (__v32hi) __B);
420 static __inline__ __m512i __DEFAULT_FN_ATTRS
421 _mm512_mask_sub_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) {
422 return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A,
428 static __inline__ __m512i __DEFAULT_FN_ATTRS
429 _mm512_maskz_sub_epi16 (__mmask32 __U, __m512i __A, __m512i __B) {
430 return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A,
433 _mm512_setzero_hi (),
437 static __inline__ __m512i __DEFAULT_FN_ATTRS
438 _mm512_mullo_epi16 (__m512i __A, __m512i __B) {
439 return (__m512i) ((__v32hi) __A * (__v32hi) __B);
442 static __inline__ __m512i __DEFAULT_FN_ATTRS
443 _mm512_mask_mullo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, __m512i __B) {
444 return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A,
450 static __inline__ __m512i __DEFAULT_FN_ATTRS
451 _mm512_maskz_mullo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) {
452 return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A,
455 _mm512_setzero_hi (),
459 static __inline__ __m512i __DEFAULT_FN_ATTRS
460 _mm512_mask_blend_epi8 (__mmask64 __U, __m512i __A, __m512i __W)
462 return (__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) __A,
467 static __inline__ __m512i __DEFAULT_FN_ATTRS
468 _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W)
470 return (__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) __A,
475 static __inline__ __m512i __DEFAULT_FN_ATTRS
476 _mm512_abs_epi8 (__m512i __A)
478 return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A,
479 (__v64qi) _mm512_setzero_qi (),
483 static __inline__ __m512i __DEFAULT_FN_ATTRS
484 _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A)
486 return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A,
491 static __inline__ __m512i __DEFAULT_FN_ATTRS
492 _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A)
494 return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A,
495 (__v64qi) _mm512_setzero_qi (),
499 static __inline__ __m512i __DEFAULT_FN_ATTRS
500 _mm512_abs_epi16 (__m512i __A)
502 return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A,
503 (__v32hi) _mm512_setzero_hi (),
507 static __inline__ __m512i __DEFAULT_FN_ATTRS
508 _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A)
510 return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A,
515 static __inline__ __m512i __DEFAULT_FN_ATTRS
516 _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A)
518 return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A,
519 (__v32hi) _mm512_setzero_hi (),
523 static __inline__ __m512i __DEFAULT_FN_ATTRS
524 _mm512_packs_epi32 (__m512i __A, __m512i __B)
526 return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A,
528 (__v32hi) _mm512_setzero_hi (),
532 static __inline__ __m512i __DEFAULT_FN_ATTRS
533 _mm512_maskz_packs_epi32 (__mmask32 __M, __m512i __A, __m512i __B)
535 return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A,
537 (__v32hi) _mm512_setzero_hi(),
541 static __inline__ __m512i __DEFAULT_FN_ATTRS
542 _mm512_mask_packs_epi32 (__m512i __W, __mmask32 __M, __m512i __A,
545 return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A,
551 static __inline__ __m512i __DEFAULT_FN_ATTRS
552 _mm512_packs_epi16 (__m512i __A, __m512i __B)
554 return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A,
556 (__v64qi) _mm512_setzero_qi (),
560 static __inline__ __m512i __DEFAULT_FN_ATTRS
561 _mm512_mask_packs_epi16 (__m512i __W, __mmask64 __M, __m512i __A,
564 return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A,
570 static __inline__ __m512i __DEFAULT_FN_ATTRS
571 _mm512_maskz_packs_epi16 (__mmask64 __M, __m512i __A, __m512i __B)
573 return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A,
575 (__v64qi) _mm512_setzero_qi(),
579 static __inline__ __m512i __DEFAULT_FN_ATTRS
580 _mm512_packus_epi32 (__m512i __A, __m512i __B)
582 return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A,
584 (__v32hi) _mm512_setzero_hi (),
588 static __inline__ __m512i __DEFAULT_FN_ATTRS
589 _mm512_maskz_packus_epi32 (__mmask32 __M, __m512i __A, __m512i __B)
591 return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A,
593 (__v32hi) _mm512_setzero_hi(),
597 static __inline__ __m512i __DEFAULT_FN_ATTRS
598 _mm512_mask_packus_epi32 (__m512i __W, __mmask32 __M, __m512i __A,
601 return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A,
607 static __inline__ __m512i __DEFAULT_FN_ATTRS
608 _mm512_packus_epi16 (__m512i __A, __m512i __B)
610 return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A,
612 (__v64qi) _mm512_setzero_qi (),
616 static __inline__ __m512i __DEFAULT_FN_ATTRS
617 _mm512_mask_packus_epi16 (__m512i __W, __mmask64 __M, __m512i __A,
620 return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A,
626 static __inline__ __m512i __DEFAULT_FN_ATTRS
627 _mm512_maskz_packus_epi16 (__mmask64 __M, __m512i __A, __m512i __B)
629 return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A,
631 (__v64qi) _mm512_setzero_qi(),
635 static __inline__ __m512i __DEFAULT_FN_ATTRS
636 _mm512_adds_epi8 (__m512i __A, __m512i __B)
638 return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A,
640 (__v64qi) _mm512_setzero_qi (),
644 static __inline__ __m512i __DEFAULT_FN_ATTRS
645 _mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A,
648 return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A,
654 static __inline__ __m512i __DEFAULT_FN_ATTRS
655 _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B)
657 return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A,
659 (__v64qi) _mm512_setzero_qi (),
663 static __inline__ __m512i __DEFAULT_FN_ATTRS
664 _mm512_adds_epi16 (__m512i __A, __m512i __B)
666 return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A,
668 (__v32hi) _mm512_setzero_hi (),
672 static __inline__ __m512i __DEFAULT_FN_ATTRS
673 _mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A,
676 return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A,
682 static __inline__ __m512i __DEFAULT_FN_ATTRS
683 _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B)
685 return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A,
687 (__v32hi) _mm512_setzero_hi (),
691 static __inline__ __m512i __DEFAULT_FN_ATTRS
692 _mm512_adds_epu8 (__m512i __A, __m512i __B)
694 return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A,
696 (__v64qi) _mm512_setzero_qi (),
700 static __inline__ __m512i __DEFAULT_FN_ATTRS
701 _mm512_mask_adds_epu8 (__m512i __W, __mmask64 __U, __m512i __A,
704 return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A,
710 static __inline__ __m512i __DEFAULT_FN_ATTRS
711 _mm512_maskz_adds_epu8 (__mmask64 __U, __m512i __A, __m512i __B)
713 return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A,
715 (__v64qi) _mm512_setzero_qi (),
719 static __inline__ __m512i __DEFAULT_FN_ATTRS
720 _mm512_adds_epu16 (__m512i __A, __m512i __B)
722 return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A,
724 (__v32hi) _mm512_setzero_hi (),
728 static __inline__ __m512i __DEFAULT_FN_ATTRS
729 _mm512_mask_adds_epu16 (__m512i __W, __mmask32 __U, __m512i __A,
732 return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A,
738 static __inline__ __m512i __DEFAULT_FN_ATTRS
739 _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
741 return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A,
743 (__v32hi) _mm512_setzero_hi (),
747 static __inline__ __m512i __DEFAULT_FN_ATTRS
748 _mm512_avg_epu8 (__m512i __A, __m512i __B)
750 return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
752 (__v64qi) _mm512_setzero_qi (),
756 static __inline__ __m512i __DEFAULT_FN_ATTRS
757 _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A,
760 return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
766 static __inline__ __m512i __DEFAULT_FN_ATTRS
767 _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B)
769 return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A,
771 (__v64qi) _mm512_setzero_qi(),
775 static __inline__ __m512i __DEFAULT_FN_ATTRS
776 _mm512_avg_epu16 (__m512i __A, __m512i __B)
778 return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
780 (__v32hi) _mm512_setzero_hi (),
784 static __inline__ __m512i __DEFAULT_FN_ATTRS
785 _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A,
788 return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
794 static __inline__ __m512i __DEFAULT_FN_ATTRS
795 _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
797 return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A,
799 (__v32hi) _mm512_setzero_hi(),
803 static __inline__ __m512i __DEFAULT_FN_ATTRS
804 _mm512_max_epi8 (__m512i __A, __m512i __B)
806 return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A,
808 (__v64qi) _mm512_setzero_qi (),
812 static __inline__ __m512i __DEFAULT_FN_ATTRS
813 _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B)
815 return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A,
817 (__v64qi) _mm512_setzero_qi(),
821 static __inline__ __m512i __DEFAULT_FN_ATTRS
822 _mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A,
825 return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A,
831 static __inline__ __m512i __DEFAULT_FN_ATTRS
832 _mm512_max_epi16 (__m512i __A, __m512i __B)
834 return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A,
836 (__v32hi) _mm512_setzero_hi (),
840 static __inline__ __m512i __DEFAULT_FN_ATTRS
841 _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B)
843 return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A,
845 (__v32hi) _mm512_setzero_hi(),
849 static __inline__ __m512i __DEFAULT_FN_ATTRS
850 _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A,
853 return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A,
859 static __inline__ __m512i __DEFAULT_FN_ATTRS
860 _mm512_max_epu8 (__m512i __A, __m512i __B)
862 return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A,
864 (__v64qi) _mm512_setzero_qi (),
868 static __inline__ __m512i __DEFAULT_FN_ATTRS
869 _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B)
871 return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A,
873 (__v64qi) _mm512_setzero_qi(),
877 static __inline__ __m512i __DEFAULT_FN_ATTRS
878 _mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A,
881 return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A,
887 static __inline__ __m512i __DEFAULT_FN_ATTRS
888 _mm512_max_epu16 (__m512i __A, __m512i __B)
890 return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A,
892 (__v32hi) _mm512_setzero_hi (),
896 static __inline__ __m512i __DEFAULT_FN_ATTRS
897 _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B)
899 return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A,
901 (__v32hi) _mm512_setzero_hi(),
905 static __inline__ __m512i __DEFAULT_FN_ATTRS
906 _mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A,
909 return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A,
915 static __inline__ __m512i __DEFAULT_FN_ATTRS
916 _mm512_min_epi8 (__m512i __A, __m512i __B)
918 return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A,
920 (__v64qi) _mm512_setzero_qi (),
924 static __inline__ __m512i __DEFAULT_FN_ATTRS
925 _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B)
927 return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A,
929 (__v64qi) _mm512_setzero_qi(),
933 static __inline__ __m512i __DEFAULT_FN_ATTRS
934 _mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A,
937 return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A,
943 static __inline__ __m512i __DEFAULT_FN_ATTRS
944 _mm512_min_epi16 (__m512i __A, __m512i __B)
946 return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A,
948 (__v32hi) _mm512_setzero_hi (),
952 static __inline__ __m512i __DEFAULT_FN_ATTRS
953 _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B)
955 return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A,
957 (__v32hi) _mm512_setzero_hi(),
961 static __inline__ __m512i __DEFAULT_FN_ATTRS
962 _mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A,
965 return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A,
971 static __inline__ __m512i __DEFAULT_FN_ATTRS
972 _mm512_min_epu8 (__m512i __A, __m512i __B)
974 return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A,
976 (__v64qi) _mm512_setzero_qi (),
980 static __inline__ __m512i __DEFAULT_FN_ATTRS
981 _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B)
983 return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A,
985 (__v64qi) _mm512_setzero_qi(),
989 static __inline__ __m512i __DEFAULT_FN_ATTRS
990 _mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A,
993 return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A,
999 static __inline__ __m512i __DEFAULT_FN_ATTRS
1000 _mm512_min_epu16 (__m512i __A, __m512i __B)
1002 return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A,
1004 (__v32hi) _mm512_setzero_hi (),
1008 static __inline__ __m512i __DEFAULT_FN_ATTRS
1009 _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B)
1011 return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A,
1013 (__v32hi) _mm512_setzero_hi(),
1017 static __inline__ __m512i __DEFAULT_FN_ATTRS
1018 _mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A,
1021 return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A,
1027 static __inline__ __m512i __DEFAULT_FN_ATTRS
1028 _mm512_shuffle_epi8 (__m512i __A, __m512i __B)
1030 return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A,
1032 (__v64qi) _mm512_setzero_qi (),
1036 static __inline__ __m512i __DEFAULT_FN_ATTRS
1037 _mm512_mask_shuffle_epi8 (__m512i __W, __mmask64 __U, __m512i __A,
1040 return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A,
1046 static __inline__ __m512i __DEFAULT_FN_ATTRS
1047 _mm512_maskz_shuffle_epi8 (__mmask64 __U, __m512i __A, __m512i __B)
1049 return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A,
1051 (__v64qi) _mm512_setzero_qi (),
1055 static __inline__ __m512i __DEFAULT_FN_ATTRS
1056 _mm512_subs_epi8 (__m512i __A, __m512i __B)
1058 return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A,
1060 (__v64qi) _mm512_setzero_qi (),
1064 static __inline__ __m512i __DEFAULT_FN_ATTRS
1065 _mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A,
1068 return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A,
1074 static __inline__ __m512i __DEFAULT_FN_ATTRS
1075 _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B)
1077 return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A,
1079 (__v64qi) _mm512_setzero_qi (),
1083 static __inline__ __m512i __DEFAULT_FN_ATTRS
1084 _mm512_subs_epi16 (__m512i __A, __m512i __B)
1086 return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A,
1088 (__v32hi) _mm512_setzero_hi (),
1092 static __inline__ __m512i __DEFAULT_FN_ATTRS
1093 _mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A,
1096 return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A,
1102 static __inline__ __m512i __DEFAULT_FN_ATTRS
1103 _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B)
1105 return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A,
1107 (__v32hi) _mm512_setzero_hi (),
1111 static __inline__ __m512i __DEFAULT_FN_ATTRS
1112 _mm512_subs_epu8 (__m512i __A, __m512i __B)
1114 return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A,
1116 (__v64qi) _mm512_setzero_qi (),
1120 static __inline__ __m512i __DEFAULT_FN_ATTRS
1121 _mm512_mask_subs_epu8 (__m512i __W, __mmask64 __U, __m512i __A,
1124 return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A,
1130 static __inline__ __m512i __DEFAULT_FN_ATTRS
1131 _mm512_maskz_subs_epu8 (__mmask64 __U, __m512i __A, __m512i __B)
1133 return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A,
1135 (__v64qi) _mm512_setzero_qi (),
1139 static __inline__ __m512i __DEFAULT_FN_ATTRS
1140 _mm512_subs_epu16 (__m512i __A, __m512i __B)
1142 return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A,
1144 (__v32hi) _mm512_setzero_hi (),
1148 static __inline__ __m512i __DEFAULT_FN_ATTRS
1149 _mm512_mask_subs_epu16 (__m512i __W, __mmask32 __U, __m512i __A,
1152 return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A,
1158 static __inline__ __m512i __DEFAULT_FN_ATTRS
1159 _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B)
1161 return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A,
1163 (__v32hi) _mm512_setzero_hi (),
1167 static __inline__ __m512i __DEFAULT_FN_ATTRS
1168 _mm512_mask2_permutex2var_epi16 (__m512i __A, __m512i __I,
1169 __mmask32 __U, __m512i __B)
1171 return (__m512i) __builtin_ia32_vpermi2varhi512_mask ((__v32hi) __A,
1172 (__v32hi) __I /* idx */ ,
1177 static __inline__ __m512i __DEFAULT_FN_ATTRS
1178 _mm512_permutex2var_epi16 (__m512i __A, __m512i __I, __m512i __B)
1180 return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I /* idx */,
1186 static __inline__ __m512i __DEFAULT_FN_ATTRS
1187 _mm512_mask_permutex2var_epi16 (__m512i __A, __mmask32 __U,
1188 __m512i __I, __m512i __B)
1190 return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I /* idx */,
1196 static __inline__ __m512i __DEFAULT_FN_ATTRS
1197 _mm512_maskz_permutex2var_epi16 (__mmask32 __U, __m512i __A,
1198 __m512i __I, __m512i __B)
1200 return (__m512i) __builtin_ia32_vpermt2varhi512_maskz ((__v32hi) __I
1207 #define _mm512_cmp_epi8_mask(a, b, p) __extension__ ({ \
1208 (__mmask16)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \
1209 (__v64qi)(__m512i)(b), \
1210 (p), (__mmask64)-1); })
1212 #define _mm512_mask_cmp_epi8_mask(m, a, b, p) __extension__ ({ \
1213 (__mmask16)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \
1214 (__v64qi)(__m512i)(b), \
1215 (p), (__mmask64)(m)); })
1217 #define _mm512_cmp_epu8_mask(a, b, p) __extension__ ({ \
1218 (__mmask16)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \
1219 (__v64qi)(__m512i)(b), \
1220 (p), (__mmask64)-1); })
1222 #define _mm512_mask_cmp_epu8_mask(m, a, b, p) __extension__ ({ \
1223 (__mmask16)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \
1224 (__v64qi)(__m512i)(b), \
1225 (p), (__mmask64)(m)); })
1227 #define _mm512_cmp_epi16_mask(a, b, p) __extension__ ({ \
1228 (__mmask16)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \
1229 (__v32hi)(__m512i)(b), \
1230 (p), (__mmask32)-1); })
1232 #define _mm512_mask_cmp_epi16_mask(m, a, b, p) __extension__ ({ \
1233 (__mmask16)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \
1234 (__v32hi)(__m512i)(b), \
1235 (p), (__mmask32)(m)); })
1237 #define _mm512_cmp_epu16_mask(a, b, p) __extension__ ({ \
1238 (__mmask16)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \
1239 (__v32hi)(__m512i)(b), \
1240 (p), (__mmask32)-1); })
1242 #define _mm512_mask_cmp_epu16_mask(m, a, b, p) __extension__ ({ \
1243 (__mmask16)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \
1244 (__v32hi)(__m512i)(b), \
1245 (p), (__mmask32)(m)); })
1248 #undef __DEFAULT_FN_ATTRS