11#error "Never use <avx512bwintrin.h> directly; include <immintrin.h> instead."
14#ifndef __AVX512BWINTRIN_H
15#define __AVX512BWINTRIN_H
21#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512bw,evex512"), __min_vector_width__(512)))
22#define __DEFAULT_FN_ATTRS64 __attribute__((__always_inline__, __nodebug__, __target__("avx512bw,evex512")))
23#define __DEFAULT_FN_ATTRS \
24 __attribute__((__always_inline__, __nodebug__, \
25 __target__("avx512bw,no-evex512")))
30 return __builtin_ia32_knotsi(__M);
36 return __builtin_ia32_knotdi(__M);
102 return (
unsigned char)__builtin_ia32_kortestcsi(__A, __B);
108 return (
unsigned char)__builtin_ia32_kortestzsi(__A, __B);
113 *__C = (
unsigned char)__builtin_ia32_kortestcsi(__A, __B);
114 return (
unsigned char)__builtin_ia32_kortestzsi(__A, __B);
120 return (
unsigned char)__builtin_ia32_kortestcdi(__A, __B);
126 return (
unsigned char)__builtin_ia32_kortestzdi(__A, __B);
131 *__C = (
unsigned char)__builtin_ia32_kortestcdi(__A, __B);
132 return (
unsigned char)__builtin_ia32_kortestzdi(__A, __B);
138 return (
unsigned char)__builtin_ia32_ktestcsi(__A, __B);
144 return (
unsigned char)__builtin_ia32_ktestzsi(__A, __B);
149 *__C = (
unsigned char)__builtin_ia32_ktestcsi(__A, __B);
150 return (
unsigned char)__builtin_ia32_ktestzsi(__A, __B);
156 return (
unsigned char)__builtin_ia32_ktestcdi(__A, __B);
162 return (
unsigned char)__builtin_ia32_ktestzdi(__A, __B);
167 *__C = (
unsigned char)__builtin_ia32_ktestcdi(__A, __B);
168 return (
unsigned char)__builtin_ia32_ktestzdi(__A, __B);
183#define _kshiftli_mask32(A, I) \
184 ((__mmask32)__builtin_ia32_kshiftlisi((__mmask32)(A), (unsigned int)(I)))
186#define _kshiftri_mask32(A, I) \
187 ((__mmask32)__builtin_ia32_kshiftrisi((__mmask32)(A), (unsigned int)(I)))
189#define _kshiftli_mask64(A, I) \
190 ((__mmask64)__builtin_ia32_kshiftlidi((__mmask64)(A), (unsigned int)(I)))
192#define _kshiftri_mask64(A, I) \
193 ((__mmask64)__builtin_ia32_kshiftridi((__mmask64)(A), (unsigned int)(I)))
197 return (
unsigned int)__builtin_ia32_kmovd((
__mmask32)__A);
202 return (
unsigned long long)__builtin_ia32_kmovq((
__mmask64)__A);
237#define _mm512_cmp_epi8_mask(a, b, p) \
238 ((__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \
239 (__v64qi)(__m512i)(b), (int)(p), \
242#define _mm512_mask_cmp_epi8_mask(m, a, b, p) \
243 ((__mmask64)__builtin_ia32_cmpb512_mask((__v64qi)(__m512i)(a), \
244 (__v64qi)(__m512i)(b), (int)(p), \
247#define _mm512_cmp_epu8_mask(a, b, p) \
248 ((__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \
249 (__v64qi)(__m512i)(b), (int)(p), \
252#define _mm512_mask_cmp_epu8_mask(m, a, b, p) \
253 ((__mmask64)__builtin_ia32_ucmpb512_mask((__v64qi)(__m512i)(a), \
254 (__v64qi)(__m512i)(b), (int)(p), \
257#define _mm512_cmp_epi16_mask(a, b, p) \
258 ((__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \
259 (__v32hi)(__m512i)(b), (int)(p), \
262#define _mm512_mask_cmp_epi16_mask(m, a, b, p) \
263 ((__mmask32)__builtin_ia32_cmpw512_mask((__v32hi)(__m512i)(a), \
264 (__v32hi)(__m512i)(b), (int)(p), \
267#define _mm512_cmp_epu16_mask(a, b, p) \
268 ((__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \
269 (__v32hi)(__m512i)(b), (int)(p), \
272#define _mm512_mask_cmp_epu16_mask(m, a, b, p) \
273 ((__mmask32)__builtin_ia32_ucmpw512_mask((__v32hi)(__m512i)(a), \
274 (__v32hi)(__m512i)(b), (int)(p), \
277#define _mm512_cmpeq_epi8_mask(A, B) \
278 _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ)
279#define _mm512_mask_cmpeq_epi8_mask(k, A, B) \
280 _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ)
281#define _mm512_cmpge_epi8_mask(A, B) \
282 _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_GE)
283#define _mm512_mask_cmpge_epi8_mask(k, A, B) \
284 _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE)
285#define _mm512_cmpgt_epi8_mask(A, B) \
286 _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_GT)
287#define _mm512_mask_cmpgt_epi8_mask(k, A, B) \
288 _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT)
289#define _mm512_cmple_epi8_mask(A, B) \
290 _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_LE)
291#define _mm512_mask_cmple_epi8_mask(k, A, B) \
292 _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE)
293#define _mm512_cmplt_epi8_mask(A, B) \
294 _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_LT)
295#define _mm512_mask_cmplt_epi8_mask(k, A, B) \
296 _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT)
297#define _mm512_cmpneq_epi8_mask(A, B) \
298 _mm512_cmp_epi8_mask((A), (B), _MM_CMPINT_NE)
299#define _mm512_mask_cmpneq_epi8_mask(k, A, B) \
300 _mm512_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE)
302#define _mm512_cmpeq_epu8_mask(A, B) \
303 _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ)
304#define _mm512_mask_cmpeq_epu8_mask(k, A, B) \
305 _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ)
306#define _mm512_cmpge_epu8_mask(A, B) \
307 _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_GE)
308#define _mm512_mask_cmpge_epu8_mask(k, A, B) \
309 _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE)
310#define _mm512_cmpgt_epu8_mask(A, B) \
311 _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_GT)
312#define _mm512_mask_cmpgt_epu8_mask(k, A, B) \
313 _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT)
314#define _mm512_cmple_epu8_mask(A, B) \
315 _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_LE)
316#define _mm512_mask_cmple_epu8_mask(k, A, B) \
317 _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE)
318#define _mm512_cmplt_epu8_mask(A, B) \
319 _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_LT)
320#define _mm512_mask_cmplt_epu8_mask(k, A, B) \
321 _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT)
322#define _mm512_cmpneq_epu8_mask(A, B) \
323 _mm512_cmp_epu8_mask((A), (B), _MM_CMPINT_NE)
324#define _mm512_mask_cmpneq_epu8_mask(k, A, B) \
325 _mm512_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE)
327#define _mm512_cmpeq_epi16_mask(A, B) \
328 _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ)
329#define _mm512_mask_cmpeq_epi16_mask(k, A, B) \
330 _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ)
331#define _mm512_cmpge_epi16_mask(A, B) \
332 _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_GE)
333#define _mm512_mask_cmpge_epi16_mask(k, A, B) \
334 _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE)
335#define _mm512_cmpgt_epi16_mask(A, B) \
336 _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_GT)
337#define _mm512_mask_cmpgt_epi16_mask(k, A, B) \
338 _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT)
339#define _mm512_cmple_epi16_mask(A, B) \
340 _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_LE)
341#define _mm512_mask_cmple_epi16_mask(k, A, B) \
342 _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE)
343#define _mm512_cmplt_epi16_mask(A, B) \
344 _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_LT)
345#define _mm512_mask_cmplt_epi16_mask(k, A, B) \
346 _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT)
347#define _mm512_cmpneq_epi16_mask(A, B) \
348 _mm512_cmp_epi16_mask((A), (B), _MM_CMPINT_NE)
349#define _mm512_mask_cmpneq_epi16_mask(k, A, B) \
350 _mm512_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE)
352#define _mm512_cmpeq_epu16_mask(A, B) \
353 _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ)
354#define _mm512_mask_cmpeq_epu16_mask(k, A, B) \
355 _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ)
356#define _mm512_cmpge_epu16_mask(A, B) \
357 _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_GE)
358#define _mm512_mask_cmpge_epu16_mask(k, A, B) \
359 _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE)
360#define _mm512_cmpgt_epu16_mask(A, B) \
361 _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_GT)
362#define _mm512_mask_cmpgt_epu16_mask(k, A, B) \
363 _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT)
364#define _mm512_cmple_epu16_mask(A, B) \
365 _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_LE)
366#define _mm512_mask_cmple_epu16_mask(k, A, B) \
367 _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE)
368#define _mm512_cmplt_epu16_mask(A, B) \
369 _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_LT)
370#define _mm512_mask_cmplt_epu16_mask(k, A, B) \
371 _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT)
372#define _mm512_cmpneq_epu16_mask(A, B) \
373 _mm512_cmp_epu16_mask((A), (B), _MM_CMPINT_NE)
374#define _mm512_mask_cmpneq_epu16_mask(k, A, B) \
375 _mm512_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE)
379 return (__m512i) ((__v64qu) __A + (__v64qu) __B);
384 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
391 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
398 return (__m512i) ((__v64qu) __A - (__v64qu) __B);
403 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
410 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
417 return (__m512i) ((__v32hu) __A + (__v32hu) __B);
422 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
429 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
436 return (__m512i) ((__v32hu) __A - (__v32hu) __B);
441 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
448 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
455 return (__m512i) ((__v32hu) __A * (__v32hu) __B);
460 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
467 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
475 return (__m512i) __builtin_ia32_selectb_512 ((
__mmask64) __U,
483 return (__m512i) __builtin_ia32_selectw_512 ((
__mmask32) __U,
491 return (__m512i)__builtin_elementwise_abs((__v64qs)__A);
497 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
505 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
513 return (__m512i)__builtin_elementwise_abs((__v32hi)__A);
519 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
527 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
535 return (__m512i)__builtin_ia32_packssdw512((__v16si)__A, (__v16si)__B);
541 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
549 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
557 return (__m512i)__builtin_ia32_packsswb512((__v32hi)__A, (__v32hi) __B);
563 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
571 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
579 return (__m512i)__builtin_ia32_packusdw512((__v16si) __A, (__v16si) __B);
585 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
593 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
601 return (__m512i)__builtin_ia32_packuswb512((__v32hi) __A, (__v32hi) __B);
607 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
615 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
623 return (__m512i)__builtin_elementwise_add_sat((__v64qs)__A, (__v64qs)__B);
629 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
637 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
645 return (__m512i)__builtin_elementwise_add_sat((__v32hi)__A, (__v32hi)__B);
651 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
659 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
667 return (__m512i)__builtin_elementwise_add_sat((__v64qu) __A, (__v64qu) __B);
673 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
681 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
689 return (__m512i)__builtin_elementwise_add_sat((__v32hu) __A, (__v32hu) __B);
695 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
703 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
711 return (__m512i)__builtin_ia32_pavgb512((__v64qi)__A, (__v64qi)__B);
718 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
726 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
734 return (__m512i)__builtin_ia32_pavgw512((__v32hi)__A, (__v32hi)__B);
741 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
749 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
757 return (__m512i)__builtin_elementwise_max((__v64qs) __A, (__v64qs) __B);
763 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
771 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
779 return (__m512i)__builtin_elementwise_max((__v32hi) __A, (__v32hi) __B);
785 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
794 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
802 return (__m512i)__builtin_elementwise_max((__v64qu)__A, (__v64qu)__B);
808 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
816 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
824 return (__m512i)__builtin_elementwise_max((__v32hu)__A, (__v32hu)__B);
830 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
838 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
846 return (__m512i)__builtin_elementwise_min((__v64qs) __A, (__v64qs) __B);
852 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
860 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
868 return (__m512i)__builtin_elementwise_min((__v32hi) __A, (__v32hi) __B);
874 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
882 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
890 return (__m512i)__builtin_elementwise_min((__v64qu)__A, (__v64qu)__B);
896 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
904 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__M,
912 return (__m512i)__builtin_elementwise_min((__v32hu)__A, (__v32hu)__B);
918 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
926 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
934 return (__m512i)__builtin_ia32_pshufb512((__v64qi)__A,(__v64qi)__B);
940 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
948 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
956 return (__m512i)__builtin_elementwise_sub_sat((__v64qs)__A, (__v64qs)__B);
962 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
970 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
978 return (__m512i)__builtin_elementwise_sub_sat((__v32hi)__A, (__v32hi)__B);
984 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
992 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1000 return (__m512i)__builtin_elementwise_sub_sat((__v64qu) __A, (__v64qu) __B);
1006 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
1014 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
1022 return (__m512i)__builtin_elementwise_sub_sat((__v32hu) __A, (__v32hu) __B);
1028 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1036 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1044 return (__m512i)__builtin_ia32_vpermi2varhi512((__v32hi)__A, (__v32hi)__I,
1052 return (__m512i)__builtin_ia32_selectw_512(__U,
1061 return (__m512i)__builtin_ia32_selectw_512(__U,
1070 return (__m512i)__builtin_ia32_selectw_512(__U,
1078 return (__m512i)__builtin_ia32_pmulhrsw512((__v32hi)__A, (__v32hi)__B);
1084 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1092 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1100 return (__m512i)__builtin_ia32_pmulhw512((__v32hi) __A, (__v32hi) __B);
1107 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1115 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1123 return (__m512i)__builtin_ia32_pmulhuw512((__v32hi) __A, (__v32hi) __B);
1129 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1137 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1144 return (__m512i)__builtin_ia32_pmaddubsw512((__v64qi)__X, (__v64qi)
__Y);
1150 return (__m512i)__builtin_ia32_selectw_512((
__mmask32) __U,
1157 return (__m512i)__builtin_ia32_selectw_512((
__mmask32) __U,
1164 return (__m512i)__builtin_ia32_pmaddwd512((__v32hi)__A, (__v32hi)__B);
1169 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__U,
1176 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__U,
1183 return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A,
1190 return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A,
1197 return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A,
1204 return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A,
1211 return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A,
1218 return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A,
1225 return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A,
1232 return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A,
1239 return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A,
1247 __builtin_ia32_pmovwb512mem_mask ((__v32qi *)
__P, (__v32hi) __A, __M);
1253 __builtin_ia32_pmovswb512mem_mask ((__v32qi *)
__P, (__v32hi) __A, __M);
1259 __builtin_ia32_pmovuswb512mem_mask ((__v32qi *)
__P, (__v32hi) __A, __M);
1264 return (__m512i)__builtin_shufflevector((__v64qi)__A, (__v64qi)__B,
1266 10, 64+10, 11, 64+11,
1267 12, 64+12, 13, 64+13,
1268 14, 64+14, 15, 64+15,
1269 24, 64+24, 25, 64+25,
1270 26, 64+26, 27, 64+27,
1271 28, 64+28, 29, 64+29,
1272 30, 64+30, 31, 64+31,
1273 40, 64+40, 41, 64+41,
1274 42, 64+42, 43, 64+43,
1275 44, 64+44, 45, 64+45,
1276 46, 64+46, 47, 64+47,
1277 56, 64+56, 57, 64+57,
1278 58, 64+58, 59, 64+59,
1279 60, 64+60, 61, 64+61,
1280 62, 64+62, 63, 64+63);
1285 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
1292 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
1299 return (__m512i)__builtin_shufflevector((__v32hi)__A, (__v32hi)__B,
1302 12, 32+12, 13, 32+13,
1303 14, 32+14, 15, 32+15,
1304 20, 32+20, 21, 32+21,
1305 22, 32+22, 23, 32+23,
1306 28, 32+28, 29, 32+29,
1307 30, 32+30, 31, 32+31);
1312 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1319 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1326 return (__m512i)__builtin_shufflevector((__v64qi)__A, (__v64qi)__B,
1331 16, 64+16, 17, 64+17,
1332 18, 64+18, 19, 64+19,
1333 20, 64+20, 21, 64+21,
1334 22, 64+22, 23, 64+23,
1335 32, 64+32, 33, 64+33,
1336 34, 64+34, 35, 64+35,
1337 36, 64+36, 37, 64+37,
1338 38, 64+38, 39, 64+39,
1339 48, 64+48, 49, 64+49,
1340 50, 64+50, 51, 64+51,
1341 52, 64+52, 53, 64+53,
1342 54, 64+54, 55, 64+55);
1347 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
1354 return (__m512i)__builtin_ia32_selectb_512((
__mmask64)__U,
1361 return (__m512i)__builtin_shufflevector((__v32hi)__A, (__v32hi)__B,
1365 10, 32+10, 11, 32+11,
1366 16, 32+16, 17, 32+17,
1367 18, 32+18, 19, 32+19,
1368 24, 32+24, 25, 32+25,
1369 26, 32+26, 27, 32+27);
1374 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1381 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1391 return (__m512i)__builtin_convertvector((__v32qs)__A, __v32hi);
1397 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1405 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1413 return (__m512i)__builtin_convertvector((__v32qu)__A, __v32hi);
1419 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1427 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1433#define _mm512_shufflehi_epi16(A, imm) \
1434 ((__m512i)__builtin_ia32_pshufhw512((__v32hi)(__m512i)(A), (int)(imm)))
1436#define _mm512_mask_shufflehi_epi16(W, U, A, imm) \
1437 ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
1438 (__v32hi)_mm512_shufflehi_epi16((A), \
1440 (__v32hi)(__m512i)(W)))
1442#define _mm512_maskz_shufflehi_epi16(U, A, imm) \
1443 ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
1444 (__v32hi)_mm512_shufflehi_epi16((A), \
1446 (__v32hi)_mm512_setzero_si512()))
1448#define _mm512_shufflelo_epi16(A, imm) \
1449 ((__m512i)__builtin_ia32_pshuflw512((__v32hi)(__m512i)(A), (int)(imm)))
1452#define _mm512_mask_shufflelo_epi16(W, U, A, imm) \
1453 ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
1454 (__v32hi)_mm512_shufflelo_epi16((A), \
1456 (__v32hi)(__m512i)(W)))
1459#define _mm512_maskz_shufflelo_epi16(U, A, imm) \
1460 ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
1461 (__v32hi)_mm512_shufflelo_epi16((A), \
1463 (__v32hi)_mm512_setzero_si512()))
1468 return (__m512i)__builtin_ia32_psllv32hi((__v32hi) __A, (__v32hi) __B);
1474 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1482 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1490 return (__m512i)__builtin_ia32_psllw512((__v32hi) __A, (__v8hi) __B);
1496 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1504 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1512 return (__m512i)__builtin_ia32_psllwi512((__v32hi)__A, (
int)__B);
1519 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1527 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1532#define _mm512_bslli_epi128(a, imm) \
1533 ((__m512i)__builtin_ia32_pslldqi512_byteshift((__v8di)(__m512i)(a), (int)(imm)))
1538 return (__m512i)__builtin_ia32_psrlv32hi((__v32hi)__A, (__v32hi)__B);
1544 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1552 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1560 return (__m512i)__builtin_ia32_psrav32hi((__v32hi)__A, (__v32hi)__B);
1566 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1574 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1582 return (__m512i)__builtin_ia32_psraw512((__v32hi) __A, (__v8hi) __B);
1588 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1596 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1604 return (__m512i)__builtin_ia32_psrawi512((__v32hi)__A, (
int)__B);
1611 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1619 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1627 return (__m512i)__builtin_ia32_psrlw512((__v32hi) __A, (__v8hi) __B);
1633 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1641 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1649 return (__m512i)__builtin_ia32_psrlwi512((__v32hi)__A, (
int)__B);
1656 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1664 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__U,
1669#define _mm512_bsrli_epi128(a, imm) \
1670 ((__m512i)__builtin_ia32_psrldqi512_byteshift((__v8di)(__m512i)(a), (int)(imm)))
1675 return (__m512i) __builtin_ia32_selectw_512 ((
__mmask32) __U,
1683 return (__m512i) __builtin_ia32_selectw_512 ((
__mmask32) __U,
1691 return (__m512i) __builtin_ia32_selectb_512 ((
__mmask64) __U,
1699 return (__m512i) __builtin_ia32_selectb_512 ((
__mmask64) __U,
1707 return (__m512i) __builtin_ia32_selectb_512(__M,
1715 return (__m512i) __builtin_ia32_selectb_512(__M,
1737 struct __loadu_epi16 {
1740 return ((
const struct __loadu_epi16*)
__P)->__v;
1746 return (__m512i) __builtin_ia32_loaddquhi512_mask ((
const __v32hi *)
__P,
1754 return (__m512i) __builtin_ia32_loaddquhi512_mask ((
const __v32hi *)
__P,
1763 struct __loadu_epi8 {
1766 return ((
const struct __loadu_epi8*)
__P)->__v;
1772 return (__m512i) __builtin_ia32_loaddquqi512_mask ((
const __v64qi *)
__P,
1780 return (__m512i) __builtin_ia32_loaddquqi512_mask ((
const __v64qi *)
__P,
1789 struct __storeu_epi16 {
1792 ((
struct __storeu_epi16*)
__P)->__v = __A;
1798 __builtin_ia32_storedquhi512_mask ((__v32hi *)
__P,
1806 struct __storeu_epi8 {
1809 ((
struct __storeu_epi8*)
__P)->__v = __A;
1815 __builtin_ia32_storedquqi512_mask ((__v64qi *)
__P,
1878 return (
__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A);
1884 return (
__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A);
1890 return (__m512i) __builtin_ia32_cvtmask2b512 (__A);
1896 return (__m512i) __builtin_ia32_cvtmask2w512 (__A);
1902 return (__m512i)__builtin_shufflevector((__v16qi) __A, (__v16qi) __A,
1903 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1904 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1905 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1906 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
1912 return (__m512i)__builtin_ia32_selectb_512(__M,
1920 return (__m512i)__builtin_ia32_selectb_512(__M,
1928 return (__m512i) __builtin_ia32_selectw_512(__M,
1936 return (__m512i) __builtin_ia32_selectw_512(__M,
1944 return (__m512i)__builtin_shufflevector((__v8hi) __A, (__v8hi) __A,
1945 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
1946 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0);
1952 return (__m512i)__builtin_ia32_selectw_512(__M,
1960 return (__m512i)__builtin_ia32_selectw_512(__M,
1968 return (__m512i)__builtin_ia32_permvarhi512((__v32hi)__B, (__v32hi)__A);
1975 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
1984 return (__m512i)__builtin_ia32_selectw_512((
__mmask32)__M,
1989#define _mm512_alignr_epi8(A, B, N) \
1990 ((__m512i)__builtin_ia32_palignr512((__v64qi)(__m512i)(A), \
1991 (__v64qi)(__m512i)(B), (int)(N)))
1993#define _mm512_mask_alignr_epi8(W, U, A, B, N) \
1994 ((__m512i)__builtin_ia32_selectb_512((__mmask64)(U), \
1995 (__v64qi)_mm512_alignr_epi8((A), (B), (int)(N)), \
1996 (__v64qi)(__m512i)(W)))
1998#define _mm512_maskz_alignr_epi8(U, A, B, N) \
1999 ((__m512i)__builtin_ia32_selectb_512((__mmask64)(U), \
2000 (__v64qi)_mm512_alignr_epi8((A), (B), (int)(N)), \
2001 (__v64qi)(__m512i)_mm512_setzero_si512()))
2003#define _mm512_dbsad_epu8(A, B, imm) \
2004 ((__m512i)__builtin_ia32_dbpsadbw512((__v64qi)(__m512i)(A), \
2005 (__v64qi)(__m512i)(B), (int)(imm)))
2007#define _mm512_mask_dbsad_epu8(W, U, A, B, imm) \
2008 ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
2009 (__v32hi)_mm512_dbsad_epu8((A), (B), (imm)), \
2010 (__v32hi)(__m512i)(W)))
2012#define _mm512_maskz_dbsad_epu8(U, A, B, imm) \
2013 ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \
2014 (__v32hi)_mm512_dbsad_epu8((A), (B), (imm)), \
2015 (__v32hi)_mm512_setzero_si512()))
2020 return (__m512i) __builtin_ia32_psadbw512 ((__v64qi) __A,
2024#undef __DEFAULT_FN_ATTRS512
2025#undef __DEFAULT_FN_ATTRS
_Float16 __2f16 __attribute__((ext_vector_type(2)))
Zeroes the upper 128 bits (bits 255:128) of all YMM registers.
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_unpackhi_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_sll_epi16(__mmask32 __U, __m512i __A, __m128i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_max_epu16(__mmask32 __M, __m512i __A, __m512i __B)
#define _mm512_cmpeq_epi16_mask(A, B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_srlv_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_subs_epu8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_permutex2var_epi16(__mmask32 __U, __m512i __A, __m512i __I, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS64 _ktestz_mask64_u8(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_madd_epi16(__mmask16 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_avg_epu8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _mm512_kunpackw(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_adds_epu16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_adds_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_unpackhi_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_subs_epu16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_sub_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _kor_mask64(__mmask64 __A, __mmask64 __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _cvtu64_mask64(unsigned long long __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_adds_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS64 _ktestc_mask64_u8(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvtepi8_epi16(__m256i __A)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _load_mask32(__mmask32 *__A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mullo_epi16(__m512i __A, __m512i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepi16_epi8(__m256i __O, __mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_broadcastw_epi16(__m512i __O, __mmask32 __M, __m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_broadcastb_epi8(__m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_avg_epu16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_shuffle_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_broadcastw_epi16(__m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_max_epu8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_set1_epi8(__m512i __O, __mmask64 __M, char __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_permutexvar_epi16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
static __inline void __DEFAULT_FN_ATTRS512 _mm512_storeu_epi8(void *__P, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask2_permutex2var_epi16(__m512i __A, __m512i __I, __mmask32 __U, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepi8_epi16(__mmask32 __U, __m256i __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS64 _kortestz_mask64_u8(__mmask64 __A, __mmask64 __B)
#define _mm512_cmpeq_epi8_mask(A, B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _kandn_mask32(__mmask32 __A, __mmask32 __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktestc_mask32_u8(__mmask32 __A, __mmask32 __B)
static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_mask_cvtsepi16_storeu_epi8(void *__P, __mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcastw_epi16(__mmask32 __M, __m128i __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_sra_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_set1_epi16(__mmask32 __M, short __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS64 _kortestc_mask64_u8(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_min_epi16(__mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_packus_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_add_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_unpackhi_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_sub_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 _mm512_mask_test_epi8_mask(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_min_epu16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
#define __DEFAULT_FN_ATTRS512
static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 _mm512_mask_testn_epi16_mask(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtusepi16_epi8(__m256i __O, __mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_abs_epi8(__m512i __W, __mmask64 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_min_epu16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_sll_epi16(__m512i __A, __m128i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_sad_epu8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_permutexvar_epi16(__m512i __A, __m512i __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _kand_mask32(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_sub_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_min_epu8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_adds_epu8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_sllv_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_srli_epi16(__m512i __W, __mmask32 __U, __m512i __A, unsigned int __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 _mm512_mask_testn_epi8_mask(__mmask64 __U, __m512i __A, __m512i __B)
#define _mm512_cmpneq_epi8_mask(A, B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_max_epi16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_subs_epu8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_slli_epi16(__m512i __A, unsigned int __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtusepi16_epi8(__mmask32 __M, __m512i __A)
#define __DEFAULT_FN_ATTRS
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mulhrs_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_sra_epi16(__m512i __A, __m128i __B)
static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_mask_storeu_epi8(void *__P, __mmask64 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_sub_epi16(__m512i __A, __m512i __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 _mm512_test_epi16_mask(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mov_epi8(__m512i __W, __mmask64 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_abs_epi16(__m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_min_epi8(__mmask64 __M, __m512i __A, __m512i __B)
static __inline __mmask64 __DEFAULT_FN_ATTRS64 _knot_mask64(__mmask64 __M)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_add_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ unsigned long long __DEFAULT_FN_ATTRS64 _cvtmask64_u64(__mmask64 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_unpacklo_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_subs_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_mask_storeu_epi16(void *__P, __mmask32 __U, __m512i __A)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_cvtusepi16_epi8(__m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_avg_epu16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_movm_epi16(__mmask32 __A)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _kadd_mask32(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mulhrs_epi16(__mmask32 __U, __m512i __A, __m512i __B)
#define _mm512_cmpneq_epi16_mask(A, B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_sll_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_srli_epi16(__m512i __A, unsigned int __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_shuffle_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_max_epi8(__mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_permutex2var_epi16(__m512i __A, __mmask32 __U, __m512i __I, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_srlv_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_packus_epi32(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mulhi_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_movm_epi8(__mmask64 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_loadu_epi8(__mmask64 __U, void const *__P)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_min_epi16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_abs_epi8(__m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mulhi_epu16(__m512i __A, __m512i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_cvtsepi16_epi8(__m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_madd_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_set1_epi8(__mmask64 __M, char __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mulhi_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_permutexvar_epi16(__mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_add_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_adds_epi8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_sub_epi8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _mm512_kunpackd(__mmask64 __A, __mmask64 __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 _mm512_movepi8_mask(__m512i __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS64 _kortest_mask64_u8(__mmask64 __A, __mmask64 __B, unsigned char *__C)
static __inline void __DEFAULT_FN_ATTRS512 _mm512_storeu_epi16(void *__P, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_maddubs_epi16(__m512i __W, __mmask32 __U, __m512i __X, __m512i __Y)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_srai_epi16(__m512i __A, unsigned int __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_max_epi8(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_min_epu8(__mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_adds_epu8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_permutex2var_epi16(__m512i __A, __m512i __I, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_adds_epu16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_srlv_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_min_epu8(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_packs_epi32(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_packus_epi16(__mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_loadu_epi16(__m512i __W, __mmask32 __U, void const *__P)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_max_epu16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_min_epu16(__mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepi8_epi16(__m512i __W, __mmask32 __U, __m256i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_unpacklo_epi8(__mmask64 __U, __m512i __A, __m512i __B)
unsigned long long __mmask64
static __inline __m512i __DEFAULT_FN_ATTRS512 _mm512_loadu_epi8(void const *__P)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_unpackhi_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_subs_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mulhi_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_avg_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_subs_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ void __DEFAULT_FN_ATTRS _store_mask32(__mmask32 *__A, __mmask32 __B)
#define _mm512_mask_cmpeq_epi8_mask(k, A, B)
#define _mm512_mask_cmpneq_epi16_mask(k, A, B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 _mm512_testn_epi16_mask(__m512i __A, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktestz_mask32_u8(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_sllv_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maddubs_epi16(__m512i __X, __m512i __Y)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_subs_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _kxor_mask32(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mulhi_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_subs_epu8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_srav_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_srli_epi16(__mmask32 __U, __m512i __A, int __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_srl_epi16(__mmask32 __U, __m512i __A, __m128i __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _kor_mask32(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_packs_epi32(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
static __inline __m512i __DEFAULT_FN_ATTRS512 _mm512_loadu_epi16(void const *__P)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtsepi16_epi8(__m256i __O, __mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_unpackhi_epi8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mullo_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_srai_epi16(__mmask32 __U, __m512i __A, unsigned int __B)
static __inline __mmask32 __DEFAULT_FN_ATTRS _knot_mask32(__mmask32 __M)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mulhrs_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_loadu_epi8(__m512i __W, __mmask64 __U, void const *__P)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_subs_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_srl_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m128i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtsepi16_epi8(__mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_maddubs_epi16(__mmask32 __U, __m512i __X, __m512i __Y)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_adds_epu16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_packus_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_max_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_packs_epi32(__mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_unpacklo_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_broadcastb_epi8(__m512i __O, __mmask64 __M, __m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_slli_epi16(__mmask32 __U, __m512i __A, unsigned int __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 _mm512_mask_test_epi16_mask(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_avg_epu8(__m512i __A, __m512i __B)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _cvtu32_mask32(unsigned int __A)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepi16_epi8(__mmask32 __M, __m512i __A)
static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_mask_cvtusepi16_storeu_epi8(void *__P, __mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_sra_epi16(__mmask32 __U, __m512i __A, __m128i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_subs_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_max_epu16(__m512i __W, __mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_packus_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_max_epu8(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_unpackhi_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_sub_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcastb_epi8(__mmask64 __M, __m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_sllv_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m256i __DEFAULT_FN_ATTRS512 _mm512_cvtepi16_epi8(__m512i __A)
#define _mm512_mask_cmpeq_epi16_mask(k, A, B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_min_epi8(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_abs_epi8(__mmask64 __U, __m512i __A)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS _kxnor_mask32(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_set1_epi16(__m512i __O, __mmask32 __M, short __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_blend_epi16(__mmask32 __U, __m512i __A, __m512i __W)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_subs_epu16(__m512i __A, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _kortestc_mask32_u8(__mmask32 __A, __mmask32 __B)
#define _mm512_mask_cmpneq_epi8_mask(k, A, B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_subs_epi8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_shuffle_epi8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_loadu_epi16(__mmask32 __U, void const *__P)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepu8_epi16(__m512i __W, __mmask32 __U, __m256i __A)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 _mm512_testn_epi8_mask(__m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS512 _mm512_test_epi8_mask(__m512i __A, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS64 _ktest_mask64_u8(__mmask64 __A, __mmask64 __B, unsigned char *__C)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _load_mask64(__mmask64 *__A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_packs_epi16(__m512i __W, __mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_unpacklo_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_abs_epi16(__m512i __W, __mmask32 __U, __m512i __A)
static __inline__ __mmask32 __DEFAULT_FN_ATTRS512 _mm512_movepi16_mask(__m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_add_epi8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_packus_epi32(__mmask32 __M, __m512i __A, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _kortest_mask32_u8(__mmask32 __A, __mmask32 __B, unsigned char *__C)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_adds_epu8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_min_epi16(__m512i __A, __m512i __B)
static __inline__ unsigned int __DEFAULT_FN_ATTRS _cvtmask32_u32(__mmask32 __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _kortestz_mask32_u8(__mmask32 __A, __mmask32 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_adds_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_blend_epi8(__mmask64 __U, __m512i __A, __m512i __W)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_max_epi16(__mmask32 __M, __m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _kand_mask64(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_srl_epi16(__m512i __A, __m128i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_max_epu8(__mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mullo_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_unpacklo_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_adds_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _kxor_mask64(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_unpacklo_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_slli_epi16(__m512i __W, __mmask32 __U, __m512i __A, unsigned int __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_packs_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mov_epi16(__m512i __W, __mmask32 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_add_epi8(__m512i __W, __mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepu8_epi16(__mmask32 __U, __m256i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_min_epi8(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mulhi_epu16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _kandn_mask64(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_adds_epi16(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvtepu8_epi16(__m256i __A)
static __inline__ void __DEFAULT_FN_ATTRS64 _store_mask64(__mmask64 *__A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_srai_epi16(__m512i __W, __mmask32 __U, __m512i __A, unsigned int __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_add_epi16(__mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_srav_epi16(__mmask32 __U, __m512i __A, __m512i __B)
#define __DEFAULT_FN_ATTRS64
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_madd_epi16(__m512i __W, __mmask16 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_avg_epu8(__mmask64 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_max_epi16(__m512i __A, __m512i __B)
static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepi16_storeu_epi8(void *__P, __mmask32 __M, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_packs_epi16(__mmask64 __M, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mov_epi16(__mmask32 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_srav_epi16(__m512i __W, __mmask32 __U, __m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mov_epi8(__mmask64 __U, __m512i __A)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _kxnor_mask64(__mmask64 __A, __mmask64 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_abs_epi16(__mmask32 __U, __m512i __A)
static __inline__ __mmask64 __DEFAULT_FN_ATTRS64 _kadd_mask64(__mmask64 __A, __mmask64 __B)
static __inline __m512i __DEFAULT_FN_ATTRS512 _mm512_set1_epi16(short __w)
static __inline __m512i __DEFAULT_FN_ATTRS512 _mm512_set1_epi8(char __w)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_and_epi32(__m512i __a, __m512i __b)
static __inline __m512i __DEFAULT_FN_ATTRS512 _mm512_setzero_si512(void)
static __inline__ __m256i __DEFAULT_FN_ATTRS _mm256_undefined_si256(void)
Create a 256-bit integer vector with undefined values.
static __inline __m256i __DEFAULT_FN_ATTRS _mm256_setzero_si256(void)
Constructs a 256-bit integer vector initialized to zero.
struct __storeu_i16 *__P __v
__inline unsigned int unsigned int unsigned int * __P
__inline unsigned int unsigned int __Y