avx512bwintrin.h revision 1.6
1 1.6 mrg /* Copyright (C) 2014-2020 Free Software Foundation, Inc. 2 1.1 mrg 3 1.1 mrg This file is part of GCC. 4 1.1 mrg 5 1.1 mrg GCC is free software; you can redistribute it and/or modify 6 1.1 mrg it under the terms of the GNU General Public License as published by 7 1.1 mrg the Free Software Foundation; either version 3, or (at your option) 8 1.1 mrg any later version. 9 1.1 mrg 10 1.1 mrg GCC is distributed in the hope that it will be useful, 11 1.1 mrg but WITHOUT ANY WARRANTY; without even the implied warranty of 12 1.1 mrg MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 13 1.1 mrg GNU General Public License for more details. 14 1.1 mrg 15 1.1 mrg Under Section 7 of GPL version 3, you are granted additional 16 1.1 mrg permissions described in the GCC Runtime Library Exception, version 17 1.1 mrg 3.1, as published by the Free Software Foundation. 18 1.1 mrg 19 1.1 mrg You should have received a copy of the GNU General Public License and 20 1.1 mrg a copy of the GCC Runtime Library Exception along with this program; 21 1.1 mrg see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 22 1.1 mrg <http://www.gnu.org/licenses/>. */ 23 1.1 mrg 24 1.1 mrg #ifndef _IMMINTRIN_H_INCLUDED 25 1.1 mrg #error "Never use <avx512bwintrin.h> directly; include <immintrin.h> instead." 26 1.1 mrg #endif 27 1.1 mrg 28 1.1 mrg #ifndef _AVX512BWINTRIN_H_INCLUDED 29 1.1 mrg #define _AVX512BWINTRIN_H_INCLUDED 30 1.1 mrg 31 1.1 mrg #ifndef __AVX512BW__ 32 1.1 mrg #pragma GCC push_options 33 1.1 mrg #pragma GCC target("avx512bw") 34 1.1 mrg #define __DISABLE_AVX512BW__ 35 1.1 mrg #endif /* __AVX512BW__ */ 36 1.1 mrg 37 1.1 mrg /* Internal data types for implementing the intrinsics. */ 38 1.1 mrg typedef short __v32hi __attribute__ ((__vector_size__ (64))); 39 1.1 mrg typedef char __v64qi __attribute__ ((__vector_size__ (64))); 40 1.1 mrg 41 1.1 mrg typedef unsigned long long __mmask64; 42 1.1 mrg 43 1.3 mrg extern __inline unsigned char 44 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 45 1.3 mrg _ktest_mask32_u8 (__mmask32 __A, __mmask32 __B, unsigned char *__CF) 46 1.3 mrg { 47 1.3 mrg *__CF = (unsigned char) __builtin_ia32_ktestcsi (__A, __B); 48 1.3 mrg return (unsigned char) __builtin_ia32_ktestzsi (__A, __B); 49 1.3 mrg } 50 1.3 mrg 51 1.3 mrg extern __inline unsigned char 52 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 53 1.3 mrg _ktest_mask64_u8 (__mmask64 __A, __mmask64 __B, unsigned char *__CF) 54 1.3 mrg { 55 1.3 mrg *__CF = (unsigned char) __builtin_ia32_ktestcdi (__A, __B); 56 1.3 mrg return (unsigned char) __builtin_ia32_ktestzdi (__A, __B); 57 1.3 mrg } 58 1.3 mrg 59 1.3 mrg extern __inline unsigned char 60 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 61 1.3 mrg _ktestz_mask32_u8 (__mmask32 __A, __mmask32 __B) 62 1.3 mrg { 63 1.3 mrg return (unsigned char) __builtin_ia32_ktestzsi (__A, __B); 64 1.3 mrg } 65 1.3 mrg 66 1.3 mrg extern __inline unsigned char 67 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 68 1.3 mrg _ktestz_mask64_u8 (__mmask64 __A, __mmask64 __B) 69 1.3 mrg { 70 1.3 mrg return (unsigned char) __builtin_ia32_ktestzdi (__A, __B); 71 1.3 mrg } 72 1.3 mrg 73 1.3 mrg extern __inline unsigned char 74 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 75 1.3 mrg _ktestc_mask32_u8 (__mmask32 __A, __mmask32 __B) 76 1.3 mrg { 77 1.3 mrg return (unsigned char) __builtin_ia32_ktestcsi (__A, __B); 78 1.3 mrg } 79 1.3 mrg 80 1.3 mrg extern __inline unsigned char 81 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 82 1.3 mrg _ktestc_mask64_u8 (__mmask64 __A, __mmask64 __B) 83 1.3 mrg { 84 1.3 mrg return (unsigned char) __builtin_ia32_ktestcdi (__A, __B); 85 1.3 mrg } 86 1.3 mrg 87 1.3 mrg extern __inline unsigned char 88 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 89 1.3 mrg _kortest_mask32_u8 (__mmask32 __A, __mmask32 __B, unsigned char *__CF) 90 1.3 mrg { 91 1.3 mrg *__CF = (unsigned char) __builtin_ia32_kortestcsi (__A, __B); 92 1.3 mrg return (unsigned char) __builtin_ia32_kortestzsi (__A, __B); 93 1.3 mrg } 94 1.3 mrg 95 1.3 mrg extern __inline unsigned char 96 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 97 1.3 mrg _kortest_mask64_u8 (__mmask64 __A, __mmask64 __B, unsigned char *__CF) 98 1.3 mrg { 99 1.3 mrg *__CF = (unsigned char) __builtin_ia32_kortestcdi (__A, __B); 100 1.3 mrg return (unsigned char) __builtin_ia32_kortestzdi (__A, __B); 101 1.3 mrg } 102 1.3 mrg 103 1.3 mrg extern __inline unsigned char 104 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 105 1.3 mrg _kortestz_mask32_u8 (__mmask32 __A, __mmask32 __B) 106 1.3 mrg { 107 1.3 mrg return (unsigned char) __builtin_ia32_kortestzsi (__A, __B); 108 1.3 mrg } 109 1.3 mrg 110 1.3 mrg extern __inline unsigned char 111 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 112 1.3 mrg _kortestz_mask64_u8 (__mmask64 __A, __mmask64 __B) 113 1.3 mrg { 114 1.3 mrg return (unsigned char) __builtin_ia32_kortestzdi (__A, __B); 115 1.3 mrg } 116 1.3 mrg 117 1.3 mrg extern __inline unsigned char 118 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 119 1.3 mrg _kortestc_mask32_u8 (__mmask32 __A, __mmask32 __B) 120 1.3 mrg { 121 1.3 mrg return (unsigned char) __builtin_ia32_kortestcsi (__A, __B); 122 1.3 mrg } 123 1.3 mrg 124 1.3 mrg extern __inline unsigned char 125 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 126 1.3 mrg _kortestc_mask64_u8 (__mmask64 __A, __mmask64 __B) 127 1.3 mrg { 128 1.3 mrg return (unsigned char) __builtin_ia32_kortestcdi (__A, __B); 129 1.3 mrg } 130 1.3 mrg 131 1.3 mrg extern __inline __mmask32 132 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 133 1.3 mrg _kadd_mask32 (__mmask32 __A, __mmask32 __B) 134 1.3 mrg { 135 1.3 mrg return (__mmask32) __builtin_ia32_kaddsi ((__mmask32) __A, (__mmask32) __B); 136 1.3 mrg } 137 1.3 mrg 138 1.3 mrg extern __inline __mmask64 139 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 140 1.3 mrg _kadd_mask64 (__mmask64 __A, __mmask64 __B) 141 1.3 mrg { 142 1.3 mrg return (__mmask64) __builtin_ia32_kadddi ((__mmask64) __A, (__mmask64) __B); 143 1.3 mrg } 144 1.3 mrg 145 1.3 mrg extern __inline unsigned int 146 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 147 1.3 mrg _cvtmask32_u32 (__mmask32 __A) 148 1.3 mrg { 149 1.3 mrg return (unsigned int) __builtin_ia32_kmovd ((__mmask32) __A); 150 1.3 mrg } 151 1.3 mrg 152 1.3 mrg extern __inline unsigned long long 153 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 154 1.3 mrg _cvtmask64_u64 (__mmask64 __A) 155 1.3 mrg { 156 1.3 mrg return (unsigned long long) __builtin_ia32_kmovq ((__mmask64) __A); 157 1.3 mrg } 158 1.3 mrg 159 1.3 mrg extern __inline __mmask32 160 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 161 1.3 mrg _cvtu32_mask32 (unsigned int __A) 162 1.3 mrg { 163 1.3 mrg return (__mmask32) __builtin_ia32_kmovd ((__mmask32) __A); 164 1.3 mrg } 165 1.3 mrg 166 1.3 mrg extern __inline __mmask64 167 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 168 1.3 mrg _cvtu64_mask64 (unsigned long long __A) 169 1.3 mrg { 170 1.3 mrg return (__mmask64) __builtin_ia32_kmovq ((__mmask64) __A); 171 1.3 mrg } 172 1.3 mrg 173 1.3 mrg extern __inline __mmask32 174 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 175 1.3 mrg _load_mask32 (__mmask32 *__A) 176 1.3 mrg { 177 1.3 mrg return (__mmask32) __builtin_ia32_kmovd (*__A); 178 1.3 mrg } 179 1.3 mrg 180 1.3 mrg extern __inline __mmask64 181 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 182 1.3 mrg _load_mask64 (__mmask64 *__A) 183 1.3 mrg { 184 1.3 mrg return (__mmask64) __builtin_ia32_kmovq (*(__mmask64 *) __A); 185 1.3 mrg } 186 1.3 mrg 187 1.3 mrg extern __inline void 188 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 189 1.3 mrg _store_mask32 (__mmask32 *__A, __mmask32 __B) 190 1.3 mrg { 191 1.3 mrg *(__mmask32 *) __A = __builtin_ia32_kmovd (__B); 192 1.3 mrg } 193 1.3 mrg 194 1.3 mrg extern __inline void 195 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 196 1.3 mrg _store_mask64 (__mmask64 *__A, __mmask64 __B) 197 1.3 mrg { 198 1.3 mrg *(__mmask64 *) __A = __builtin_ia32_kmovq (__B); 199 1.3 mrg } 200 1.3 mrg 201 1.3 mrg extern __inline __mmask32 202 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 203 1.3 mrg _knot_mask32 (__mmask32 __A) 204 1.3 mrg { 205 1.3 mrg return (__mmask32) __builtin_ia32_knotsi ((__mmask32) __A); 206 1.3 mrg } 207 1.3 mrg 208 1.3 mrg extern __inline __mmask64 209 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 210 1.3 mrg _knot_mask64 (__mmask64 __A) 211 1.3 mrg { 212 1.3 mrg return (__mmask64) __builtin_ia32_knotdi ((__mmask64) __A); 213 1.3 mrg } 214 1.3 mrg 215 1.3 mrg extern __inline __mmask32 216 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 217 1.3 mrg _kor_mask32 (__mmask32 __A, __mmask32 __B) 218 1.3 mrg { 219 1.3 mrg return (__mmask32) __builtin_ia32_korsi ((__mmask32) __A, (__mmask32) __B); 220 1.3 mrg } 221 1.3 mrg 222 1.3 mrg extern __inline __mmask64 223 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 224 1.3 mrg _kor_mask64 (__mmask64 __A, __mmask64 __B) 225 1.3 mrg { 226 1.3 mrg return (__mmask64) __builtin_ia32_kordi ((__mmask64) __A, (__mmask64) __B); 227 1.3 mrg } 228 1.3 mrg 229 1.3 mrg extern __inline __mmask32 230 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 231 1.3 mrg _kxnor_mask32 (__mmask32 __A, __mmask32 __B) 232 1.3 mrg { 233 1.3 mrg return (__mmask32) __builtin_ia32_kxnorsi ((__mmask32) __A, (__mmask32) __B); 234 1.3 mrg } 235 1.3 mrg 236 1.3 mrg extern __inline __mmask64 237 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 238 1.3 mrg _kxnor_mask64 (__mmask64 __A, __mmask64 __B) 239 1.3 mrg { 240 1.3 mrg return (__mmask64) __builtin_ia32_kxnordi ((__mmask64) __A, (__mmask64) __B); 241 1.3 mrg } 242 1.3 mrg 243 1.3 mrg extern __inline __mmask32 244 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 245 1.3 mrg _kxor_mask32 (__mmask32 __A, __mmask32 __B) 246 1.3 mrg { 247 1.3 mrg return (__mmask32) __builtin_ia32_kxorsi ((__mmask32) __A, (__mmask32) __B); 248 1.3 mrg } 249 1.3 mrg 250 1.3 mrg extern __inline __mmask64 251 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 252 1.3 mrg _kxor_mask64 (__mmask64 __A, __mmask64 __B) 253 1.3 mrg { 254 1.3 mrg return (__mmask64) __builtin_ia32_kxordi ((__mmask64) __A, (__mmask64) __B); 255 1.3 mrg } 256 1.3 mrg 257 1.3 mrg extern __inline __mmask32 258 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 259 1.3 mrg _kand_mask32 (__mmask32 __A, __mmask32 __B) 260 1.3 mrg { 261 1.3 mrg return (__mmask32) __builtin_ia32_kandsi ((__mmask32) __A, (__mmask32) __B); 262 1.3 mrg } 263 1.3 mrg 264 1.3 mrg extern __inline __mmask64 265 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 266 1.3 mrg _kand_mask64 (__mmask64 __A, __mmask64 __B) 267 1.3 mrg { 268 1.3 mrg return (__mmask64) __builtin_ia32_kanddi ((__mmask64) __A, (__mmask64) __B); 269 1.3 mrg } 270 1.3 mrg 271 1.3 mrg extern __inline __mmask32 272 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 273 1.3 mrg _kandn_mask32 (__mmask32 __A, __mmask32 __B) 274 1.1 mrg { 275 1.3 mrg return (__mmask32) __builtin_ia32_kandnsi ((__mmask32) __A, (__mmask32) __B); 276 1.1 mrg } 277 1.1 mrg 278 1.3 mrg extern __inline __mmask64 279 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 280 1.3 mrg _kandn_mask64 (__mmask64 __A, __mmask64 __B) 281 1.1 mrg { 282 1.3 mrg return (__mmask64) __builtin_ia32_kandndi ((__mmask64) __A, (__mmask64) __B); 283 1.1 mrg } 284 1.1 mrg 285 1.1 mrg extern __inline __m512i 286 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 287 1.1 mrg _mm512_mask_mov_epi16 (__m512i __W, __mmask32 __U, __m512i __A) 288 1.1 mrg { 289 1.1 mrg return (__m512i) __builtin_ia32_movdquhi512_mask ((__v32hi) __A, 290 1.1 mrg (__v32hi) __W, 291 1.1 mrg (__mmask32) __U); 292 1.1 mrg } 293 1.1 mrg 294 1.1 mrg extern __inline __m512i 295 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 296 1.1 mrg _mm512_maskz_mov_epi16 (__mmask32 __U, __m512i __A) 297 1.1 mrg { 298 1.1 mrg return (__m512i) __builtin_ia32_movdquhi512_mask ((__v32hi) __A, 299 1.1 mrg (__v32hi) 300 1.3 mrg _mm512_setzero_si512 (), 301 1.1 mrg (__mmask32) __U); 302 1.1 mrg } 303 1.1 mrg 304 1.1 mrg extern __inline __m512i 305 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 306 1.1 mrg _mm512_mask_loadu_epi16 (__m512i __W, __mmask32 __U, void const *__P) 307 1.1 mrg { 308 1.3 mrg return (__m512i) __builtin_ia32_loaddquhi512_mask ((const short *) __P, 309 1.1 mrg (__v32hi) __W, 310 1.1 mrg (__mmask32) __U); 311 1.1 mrg } 312 1.1 mrg 313 1.1 mrg extern __inline __m512i 314 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 315 1.1 mrg _mm512_maskz_loadu_epi16 (__mmask32 __U, void const *__P) 316 1.1 mrg { 317 1.3 mrg return (__m512i) __builtin_ia32_loaddquhi512_mask ((const short *) __P, 318 1.1 mrg (__v32hi) 319 1.3 mrg _mm512_setzero_si512 (), 320 1.1 mrg (__mmask32) __U); 321 1.1 mrg } 322 1.1 mrg 323 1.1 mrg extern __inline void 324 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 325 1.1 mrg _mm512_mask_storeu_epi16 (void *__P, __mmask32 __U, __m512i __A) 326 1.1 mrg { 327 1.3 mrg __builtin_ia32_storedquhi512_mask ((short *) __P, 328 1.1 mrg (__v32hi) __A, 329 1.1 mrg (__mmask32) __U); 330 1.1 mrg } 331 1.1 mrg 332 1.1 mrg extern __inline __m512i 333 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 334 1.1 mrg _mm512_mask_mov_epi8 (__m512i __W, __mmask64 __U, __m512i __A) 335 1.1 mrg { 336 1.1 mrg return (__m512i) __builtin_ia32_movdquqi512_mask ((__v64qi) __A, 337 1.1 mrg (__v64qi) __W, 338 1.1 mrg (__mmask64) __U); 339 1.1 mrg } 340 1.1 mrg 341 1.1 mrg extern __inline __m512i 342 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 343 1.1 mrg _mm512_maskz_mov_epi8 (__mmask64 __U, __m512i __A) 344 1.1 mrg { 345 1.1 mrg return (__m512i) __builtin_ia32_movdquqi512_mask ((__v64qi) __A, 346 1.1 mrg (__v64qi) 347 1.3 mrg _mm512_setzero_si512 (), 348 1.1 mrg (__mmask64) __U); 349 1.1 mrg } 350 1.1 mrg 351 1.1 mrg extern __inline __mmask32 352 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 353 1.1 mrg _mm512_kunpackw (__mmask32 __A, __mmask32 __B) 354 1.1 mrg { 355 1.1 mrg return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, 356 1.1 mrg (__mmask32) __B); 357 1.1 mrg } 358 1.1 mrg 359 1.3 mrg extern __inline __mmask32 360 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 361 1.3 mrg _kunpackw_mask32 (__mmask16 __A, __mmask16 __B) 362 1.3 mrg { 363 1.3 mrg return (__mmask32) __builtin_ia32_kunpcksi ((__mmask32) __A, 364 1.3 mrg (__mmask32) __B); 365 1.3 mrg } 366 1.3 mrg 367 1.1 mrg extern __inline __mmask64 368 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 369 1.1 mrg _mm512_kunpackd (__mmask64 __A, __mmask64 __B) 370 1.1 mrg { 371 1.1 mrg return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, 372 1.1 mrg (__mmask64) __B); 373 1.1 mrg } 374 1.1 mrg 375 1.3 mrg extern __inline __mmask64 376 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 377 1.3 mrg _kunpackd_mask64 (__mmask32 __A, __mmask32 __B) 378 1.3 mrg { 379 1.3 mrg return (__mmask64) __builtin_ia32_kunpckdi ((__mmask64) __A, 380 1.3 mrg (__mmask64) __B); 381 1.3 mrg } 382 1.3 mrg 383 1.1 mrg extern __inline __m512i 384 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 385 1.1 mrg _mm512_mask_loadu_epi8 (__m512i __W, __mmask64 __U, void const *__P) 386 1.1 mrg { 387 1.3 mrg return (__m512i) __builtin_ia32_loaddquqi512_mask ((const char *) __P, 388 1.1 mrg (__v64qi) __W, 389 1.1 mrg (__mmask64) __U); 390 1.1 mrg } 391 1.1 mrg 392 1.1 mrg extern __inline __m512i 393 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 394 1.1 mrg _mm512_maskz_loadu_epi8 (__mmask64 __U, void const *__P) 395 1.1 mrg { 396 1.3 mrg return (__m512i) __builtin_ia32_loaddquqi512_mask ((const char *) __P, 397 1.1 mrg (__v64qi) 398 1.3 mrg _mm512_setzero_si512 (), 399 1.1 mrg (__mmask64) __U); 400 1.1 mrg } 401 1.1 mrg 402 1.1 mrg extern __inline void 403 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 404 1.1 mrg _mm512_mask_storeu_epi8 (void *__P, __mmask64 __U, __m512i __A) 405 1.1 mrg { 406 1.3 mrg __builtin_ia32_storedquqi512_mask ((char *) __P, 407 1.1 mrg (__v64qi) __A, 408 1.1 mrg (__mmask64) __U); 409 1.1 mrg } 410 1.1 mrg 411 1.1 mrg extern __inline __m512i 412 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 413 1.1 mrg _mm512_sad_epu8 (__m512i __A, __m512i __B) 414 1.1 mrg { 415 1.1 mrg return (__m512i) __builtin_ia32_psadbw512 ((__v64qi) __A, 416 1.1 mrg (__v64qi) __B); 417 1.1 mrg } 418 1.1 mrg 419 1.1 mrg extern __inline __m256i 420 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 421 1.1 mrg _mm512_cvtepi16_epi8 (__m512i __A) 422 1.1 mrg { 423 1.1 mrg return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, 424 1.1 mrg (__v32qi) _mm256_undefined_si256(), 425 1.1 mrg (__mmask32) -1); 426 1.1 mrg } 427 1.1 mrg 428 1.4 mrg extern __inline void 429 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 430 1.4 mrg _mm512_mask_cvtepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) 431 1.4 mrg { 432 1.4 mrg __builtin_ia32_pmovwb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); 433 1.4 mrg } 434 1.4 mrg 435 1.1 mrg extern __inline __m256i 436 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 437 1.1 mrg _mm512_mask_cvtepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) 438 1.1 mrg { 439 1.1 mrg return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, 440 1.1 mrg (__v32qi) __O, __M); 441 1.1 mrg } 442 1.1 mrg 443 1.1 mrg extern __inline __m256i 444 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 445 1.1 mrg _mm512_maskz_cvtepi16_epi8 (__mmask32 __M, __m512i __A) 446 1.1 mrg { 447 1.1 mrg return (__m256i) __builtin_ia32_pmovwb512_mask ((__v32hi) __A, 448 1.1 mrg (__v32qi) 449 1.1 mrg _mm256_setzero_si256 (), 450 1.1 mrg __M); 451 1.1 mrg } 452 1.1 mrg 453 1.1 mrg extern __inline __m256i 454 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 455 1.1 mrg _mm512_cvtsepi16_epi8 (__m512i __A) 456 1.1 mrg { 457 1.1 mrg return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, 458 1.1 mrg (__v32qi)_mm256_undefined_si256(), 459 1.1 mrg (__mmask32) -1); 460 1.1 mrg } 461 1.1 mrg 462 1.4 mrg extern __inline void 463 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 464 1.4 mrg _mm512_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) 465 1.4 mrg { 466 1.4 mrg __builtin_ia32_pmovswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); 467 1.4 mrg } 468 1.4 mrg 469 1.1 mrg extern __inline __m256i 470 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 471 1.1 mrg _mm512_mask_cvtsepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) 472 1.1 mrg { 473 1.1 mrg return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, 474 1.1 mrg (__v32qi)__O, 475 1.1 mrg __M); 476 1.1 mrg } 477 1.1 mrg 478 1.1 mrg extern __inline __m256i 479 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 480 1.1 mrg _mm512_maskz_cvtsepi16_epi8 (__mmask32 __M, __m512i __A) 481 1.1 mrg { 482 1.1 mrg return (__m256i) __builtin_ia32_pmovswb512_mask ((__v32hi) __A, 483 1.1 mrg (__v32qi) 484 1.1 mrg _mm256_setzero_si256 (), 485 1.1 mrg __M); 486 1.1 mrg } 487 1.1 mrg 488 1.1 mrg extern __inline __m256i 489 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 490 1.1 mrg _mm512_cvtusepi16_epi8 (__m512i __A) 491 1.1 mrg { 492 1.1 mrg return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, 493 1.1 mrg (__v32qi)_mm256_undefined_si256(), 494 1.1 mrg (__mmask32) -1); 495 1.1 mrg } 496 1.1 mrg 497 1.1 mrg extern __inline __m256i 498 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 499 1.1 mrg _mm512_mask_cvtusepi16_epi8 (__m256i __O, __mmask32 __M, __m512i __A) 500 1.1 mrg { 501 1.1 mrg return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, 502 1.1 mrg (__v32qi) __O, 503 1.1 mrg __M); 504 1.1 mrg } 505 1.1 mrg 506 1.4 mrg extern __inline void 507 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 508 1.4 mrg _mm512_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask32 __M, __m512i __A) 509 1.4 mrg { 510 1.4 mrg __builtin_ia32_pmovuswb512mem_mask ((__v32qi *) __P, (__v32hi) __A, __M); 511 1.4 mrg } 512 1.4 mrg 513 1.1 mrg extern __inline __m256i 514 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 515 1.1 mrg _mm512_maskz_cvtusepi16_epi8 (__mmask32 __M, __m512i __A) 516 1.1 mrg { 517 1.1 mrg return (__m256i) __builtin_ia32_pmovuswb512_mask ((__v32hi) __A, 518 1.1 mrg (__v32qi) 519 1.1 mrg _mm256_setzero_si256 (), 520 1.1 mrg __M); 521 1.1 mrg } 522 1.1 mrg 523 1.1 mrg extern __inline __m512i 524 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 525 1.1 mrg _mm512_broadcastb_epi8 (__m128i __A) 526 1.1 mrg { 527 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, 528 1.3 mrg (__v64qi)_mm512_undefined_epi32(), 529 1.3 mrg (__mmask64) -1); 530 1.1 mrg } 531 1.1 mrg 532 1.1 mrg extern __inline __m512i 533 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 534 1.1 mrg _mm512_mask_broadcastb_epi8 (__m512i __O, __mmask64 __M, __m128i __A) 535 1.1 mrg { 536 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, 537 1.1 mrg (__v64qi) __O, 538 1.1 mrg __M); 539 1.1 mrg } 540 1.1 mrg 541 1.1 mrg extern __inline __m512i 542 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 543 1.1 mrg _mm512_maskz_broadcastb_epi8 (__mmask64 __M, __m128i __A) 544 1.1 mrg { 545 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastb512_mask ((__v16qi) __A, 546 1.1 mrg (__v64qi) 547 1.3 mrg _mm512_setzero_si512 (), 548 1.1 mrg __M); 549 1.1 mrg } 550 1.1 mrg 551 1.1 mrg extern __inline __m512i 552 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 553 1.1 mrg _mm512_mask_set1_epi8 (__m512i __O, __mmask64 __M, char __A) 554 1.1 mrg { 555 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastb512_gpr_mask (__A, 556 1.1 mrg (__v64qi) __O, 557 1.1 mrg __M); 558 1.1 mrg } 559 1.1 mrg 560 1.1 mrg extern __inline __m512i 561 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 562 1.1 mrg _mm512_maskz_set1_epi8 (__mmask64 __M, char __A) 563 1.1 mrg { 564 1.3 mrg return (__m512i) 565 1.3 mrg __builtin_ia32_pbroadcastb512_gpr_mask (__A, 566 1.3 mrg (__v64qi) 567 1.3 mrg _mm512_setzero_si512 (), 568 1.3 mrg __M); 569 1.1 mrg } 570 1.1 mrg 571 1.1 mrg extern __inline __m512i 572 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 573 1.1 mrg _mm512_broadcastw_epi16 (__m128i __A) 574 1.1 mrg { 575 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, 576 1.3 mrg (__v32hi)_mm512_undefined_epi32(), 577 1.3 mrg (__mmask32) -1); 578 1.1 mrg } 579 1.1 mrg 580 1.1 mrg extern __inline __m512i 581 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 582 1.1 mrg _mm512_mask_broadcastw_epi16 (__m512i __O, __mmask32 __M, __m128i __A) 583 1.1 mrg { 584 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, 585 1.1 mrg (__v32hi) __O, 586 1.1 mrg __M); 587 1.1 mrg } 588 1.1 mrg 589 1.1 mrg extern __inline __m512i 590 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 591 1.1 mrg _mm512_maskz_broadcastw_epi16 (__mmask32 __M, __m128i __A) 592 1.1 mrg { 593 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastw512_mask ((__v8hi) __A, 594 1.1 mrg (__v32hi) 595 1.3 mrg _mm512_setzero_si512 (), 596 1.1 mrg __M); 597 1.1 mrg } 598 1.1 mrg 599 1.1 mrg extern __inline __m512i 600 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 601 1.1 mrg _mm512_mask_set1_epi16 (__m512i __O, __mmask32 __M, short __A) 602 1.1 mrg { 603 1.1 mrg return (__m512i) __builtin_ia32_pbroadcastw512_gpr_mask (__A, 604 1.1 mrg (__v32hi) __O, 605 1.1 mrg __M); 606 1.1 mrg } 607 1.1 mrg 608 1.1 mrg extern __inline __m512i 609 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 610 1.1 mrg _mm512_maskz_set1_epi16 (__mmask32 __M, short __A) 611 1.1 mrg { 612 1.3 mrg return (__m512i) 613 1.3 mrg __builtin_ia32_pbroadcastw512_gpr_mask (__A, 614 1.3 mrg (__v32hi) 615 1.3 mrg _mm512_setzero_si512 (), 616 1.3 mrg __M); 617 1.1 mrg } 618 1.1 mrg 619 1.1 mrg extern __inline __m512i 620 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 621 1.1 mrg _mm512_mulhrs_epi16 (__m512i __A, __m512i __B) 622 1.1 mrg { 623 1.1 mrg return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, 624 1.1 mrg (__v32hi) __B, 625 1.1 mrg (__v32hi) 626 1.3 mrg _mm512_setzero_si512 (), 627 1.1 mrg (__mmask32) -1); 628 1.1 mrg } 629 1.1 mrg 630 1.1 mrg extern __inline __m512i 631 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 632 1.1 mrg _mm512_mask_mulhrs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 633 1.1 mrg __m512i __B) 634 1.1 mrg { 635 1.1 mrg return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, 636 1.1 mrg (__v32hi) __B, 637 1.1 mrg (__v32hi) __W, 638 1.1 mrg (__mmask32) __U); 639 1.1 mrg } 640 1.1 mrg 641 1.1 mrg extern __inline __m512i 642 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 643 1.1 mrg _mm512_maskz_mulhrs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 644 1.1 mrg { 645 1.1 mrg return (__m512i) __builtin_ia32_pmulhrsw512_mask ((__v32hi) __A, 646 1.1 mrg (__v32hi) __B, 647 1.1 mrg (__v32hi) 648 1.3 mrg _mm512_setzero_si512 (), 649 1.1 mrg (__mmask32) __U); 650 1.1 mrg } 651 1.1 mrg 652 1.1 mrg extern __inline __m512i 653 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 654 1.1 mrg _mm512_mulhi_epi16 (__m512i __A, __m512i __B) 655 1.1 mrg { 656 1.1 mrg return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, 657 1.1 mrg (__v32hi) __B, 658 1.1 mrg (__v32hi) 659 1.3 mrg _mm512_setzero_si512 (), 660 1.1 mrg (__mmask32) -1); 661 1.1 mrg } 662 1.1 mrg 663 1.1 mrg extern __inline __m512i 664 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 665 1.1 mrg _mm512_mask_mulhi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 666 1.1 mrg __m512i __B) 667 1.1 mrg { 668 1.1 mrg return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, 669 1.1 mrg (__v32hi) __B, 670 1.1 mrg (__v32hi) __W, 671 1.1 mrg (__mmask32) __U); 672 1.1 mrg } 673 1.1 mrg 674 1.1 mrg extern __inline __m512i 675 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 676 1.1 mrg _mm512_maskz_mulhi_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 677 1.1 mrg { 678 1.1 mrg return (__m512i) __builtin_ia32_pmulhw512_mask ((__v32hi) __A, 679 1.1 mrg (__v32hi) __B, 680 1.1 mrg (__v32hi) 681 1.3 mrg _mm512_setzero_si512 (), 682 1.1 mrg (__mmask32) __U); 683 1.1 mrg } 684 1.1 mrg 685 1.1 mrg extern __inline __m512i 686 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 687 1.1 mrg _mm512_mulhi_epu16 (__m512i __A, __m512i __B) 688 1.1 mrg { 689 1.1 mrg return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, 690 1.1 mrg (__v32hi) __B, 691 1.1 mrg (__v32hi) 692 1.3 mrg _mm512_setzero_si512 (), 693 1.1 mrg (__mmask32) -1); 694 1.1 mrg } 695 1.1 mrg 696 1.1 mrg extern __inline __m512i 697 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 698 1.1 mrg _mm512_mask_mulhi_epu16 (__m512i __W, __mmask32 __U, __m512i __A, 699 1.1 mrg __m512i __B) 700 1.1 mrg { 701 1.1 mrg return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, 702 1.1 mrg (__v32hi) __B, 703 1.1 mrg (__v32hi) __W, 704 1.1 mrg (__mmask32) __U); 705 1.1 mrg } 706 1.1 mrg 707 1.1 mrg extern __inline __m512i 708 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 709 1.1 mrg _mm512_maskz_mulhi_epu16 (__mmask32 __U, __m512i __A, __m512i __B) 710 1.1 mrg { 711 1.1 mrg return (__m512i) __builtin_ia32_pmulhuw512_mask ((__v32hi) __A, 712 1.1 mrg (__v32hi) __B, 713 1.1 mrg (__v32hi) 714 1.3 mrg _mm512_setzero_si512 (), 715 1.1 mrg (__mmask32) __U); 716 1.1 mrg } 717 1.1 mrg 718 1.1 mrg extern __inline __m512i 719 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 720 1.1 mrg _mm512_mullo_epi16 (__m512i __A, __m512i __B) 721 1.1 mrg { 722 1.1 mrg return (__m512i) ((__v32hu) __A * (__v32hu) __B); 723 1.1 mrg } 724 1.1 mrg 725 1.1 mrg extern __inline __m512i 726 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 727 1.1 mrg _mm512_mask_mullo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 728 1.1 mrg __m512i __B) 729 1.1 mrg { 730 1.1 mrg return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A, 731 1.1 mrg (__v32hi) __B, 732 1.1 mrg (__v32hi) __W, 733 1.1 mrg (__mmask32) __U); 734 1.1 mrg } 735 1.1 mrg 736 1.1 mrg extern __inline __m512i 737 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 738 1.1 mrg _mm512_maskz_mullo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 739 1.1 mrg { 740 1.1 mrg return (__m512i) __builtin_ia32_pmullw512_mask ((__v32hi) __A, 741 1.1 mrg (__v32hi) __B, 742 1.1 mrg (__v32hi) 743 1.3 mrg _mm512_setzero_si512 (), 744 1.1 mrg (__mmask32) __U); 745 1.1 mrg } 746 1.1 mrg 747 1.1 mrg extern __inline __m512i 748 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 749 1.1 mrg _mm512_cvtepi8_epi16 (__m256i __A) 750 1.1 mrg { 751 1.1 mrg return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, 752 1.1 mrg (__v32hi) 753 1.3 mrg _mm512_setzero_si512 (), 754 1.1 mrg (__mmask32) -1); 755 1.1 mrg } 756 1.1 mrg 757 1.1 mrg extern __inline __m512i 758 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 759 1.1 mrg _mm512_mask_cvtepi8_epi16 (__m512i __W, __mmask32 __U, __m256i __A) 760 1.1 mrg { 761 1.1 mrg return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, 762 1.1 mrg (__v32hi) __W, 763 1.1 mrg (__mmask32) __U); 764 1.1 mrg } 765 1.1 mrg 766 1.1 mrg extern __inline __m512i 767 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 768 1.1 mrg _mm512_maskz_cvtepi8_epi16 (__mmask32 __U, __m256i __A) 769 1.1 mrg { 770 1.1 mrg return (__m512i) __builtin_ia32_pmovsxbw512_mask ((__v32qi) __A, 771 1.1 mrg (__v32hi) 772 1.3 mrg _mm512_setzero_si512 (), 773 1.1 mrg (__mmask32) __U); 774 1.1 mrg } 775 1.1 mrg 776 1.1 mrg extern __inline __m512i 777 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 778 1.1 mrg _mm512_cvtepu8_epi16 (__m256i __A) 779 1.1 mrg { 780 1.1 mrg return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, 781 1.1 mrg (__v32hi) 782 1.3 mrg _mm512_setzero_si512 (), 783 1.1 mrg (__mmask32) -1); 784 1.1 mrg } 785 1.1 mrg 786 1.1 mrg extern __inline __m512i 787 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 788 1.1 mrg _mm512_mask_cvtepu8_epi16 (__m512i __W, __mmask32 __U, __m256i __A) 789 1.1 mrg { 790 1.1 mrg return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, 791 1.1 mrg (__v32hi) __W, 792 1.1 mrg (__mmask32) __U); 793 1.1 mrg } 794 1.1 mrg 795 1.1 mrg extern __inline __m512i 796 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 797 1.1 mrg _mm512_maskz_cvtepu8_epi16 (__mmask32 __U, __m256i __A) 798 1.1 mrg { 799 1.1 mrg return (__m512i) __builtin_ia32_pmovzxbw512_mask ((__v32qi) __A, 800 1.1 mrg (__v32hi) 801 1.3 mrg _mm512_setzero_si512 (), 802 1.1 mrg (__mmask32) __U); 803 1.1 mrg } 804 1.1 mrg 805 1.1 mrg extern __inline __m512i 806 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 807 1.1 mrg _mm512_permutexvar_epi16 (__m512i __A, __m512i __B) 808 1.1 mrg { 809 1.1 mrg return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, 810 1.1 mrg (__v32hi) __A, 811 1.1 mrg (__v32hi) 812 1.3 mrg _mm512_setzero_si512 (), 813 1.1 mrg (__mmask32) -1); 814 1.1 mrg } 815 1.1 mrg 816 1.1 mrg extern __inline __m512i 817 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 818 1.1 mrg _mm512_maskz_permutexvar_epi16 (__mmask32 __M, __m512i __A, 819 1.1 mrg __m512i __B) 820 1.1 mrg { 821 1.1 mrg return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, 822 1.1 mrg (__v32hi) __A, 823 1.1 mrg (__v32hi) 824 1.3 mrg _mm512_setzero_si512 (), 825 1.1 mrg (__mmask32) __M); 826 1.1 mrg } 827 1.1 mrg 828 1.1 mrg extern __inline __m512i 829 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 830 1.1 mrg _mm512_mask_permutexvar_epi16 (__m512i __W, __mmask32 __M, __m512i __A, 831 1.1 mrg __m512i __B) 832 1.1 mrg { 833 1.1 mrg return (__m512i) __builtin_ia32_permvarhi512_mask ((__v32hi) __B, 834 1.1 mrg (__v32hi) __A, 835 1.1 mrg (__v32hi) __W, 836 1.1 mrg (__mmask32) __M); 837 1.1 mrg } 838 1.1 mrg 839 1.1 mrg extern __inline __m512i 840 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 841 1.1 mrg _mm512_permutex2var_epi16 (__m512i __A, __m512i __I, __m512i __B) 842 1.1 mrg { 843 1.1 mrg return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I 844 1.1 mrg /* idx */ , 845 1.1 mrg (__v32hi) __A, 846 1.1 mrg (__v32hi) __B, 847 1.3 mrg (__mmask32) -1); 848 1.1 mrg } 849 1.1 mrg 850 1.1 mrg extern __inline __m512i 851 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 852 1.1 mrg _mm512_mask_permutex2var_epi16 (__m512i __A, __mmask32 __U, 853 1.1 mrg __m512i __I, __m512i __B) 854 1.1 mrg { 855 1.1 mrg return (__m512i) __builtin_ia32_vpermt2varhi512_mask ((__v32hi) __I 856 1.1 mrg /* idx */ , 857 1.1 mrg (__v32hi) __A, 858 1.1 mrg (__v32hi) __B, 859 1.1 mrg (__mmask32) 860 1.1 mrg __U); 861 1.1 mrg } 862 1.1 mrg 863 1.1 mrg extern __inline __m512i 864 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 865 1.1 mrg _mm512_mask2_permutex2var_epi16 (__m512i __A, __m512i __I, 866 1.1 mrg __mmask32 __U, __m512i __B) 867 1.1 mrg { 868 1.1 mrg return (__m512i) __builtin_ia32_vpermi2varhi512_mask ((__v32hi) __A, 869 1.1 mrg (__v32hi) __I 870 1.1 mrg /* idx */ , 871 1.1 mrg (__v32hi) __B, 872 1.1 mrg (__mmask32) 873 1.1 mrg __U); 874 1.1 mrg } 875 1.1 mrg 876 1.1 mrg extern __inline __m512i 877 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 878 1.1 mrg _mm512_maskz_permutex2var_epi16 (__mmask32 __U, __m512i __A, 879 1.1 mrg __m512i __I, __m512i __B) 880 1.1 mrg { 881 1.1 mrg return (__m512i) __builtin_ia32_vpermt2varhi512_maskz ((__v32hi) __I 882 1.1 mrg /* idx */ , 883 1.1 mrg (__v32hi) __A, 884 1.1 mrg (__v32hi) __B, 885 1.1 mrg (__mmask32) 886 1.1 mrg __U); 887 1.1 mrg } 888 1.1 mrg 889 1.1 mrg extern __inline __m512i 890 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 891 1.1 mrg _mm512_avg_epu8 (__m512i __A, __m512i __B) 892 1.1 mrg { 893 1.1 mrg return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, 894 1.1 mrg (__v64qi) __B, 895 1.1 mrg (__v64qi) 896 1.3 mrg _mm512_setzero_si512 (), 897 1.1 mrg (__mmask64) -1); 898 1.1 mrg } 899 1.1 mrg 900 1.1 mrg extern __inline __m512i 901 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 902 1.1 mrg _mm512_mask_avg_epu8 (__m512i __W, __mmask64 __U, __m512i __A, 903 1.1 mrg __m512i __B) 904 1.1 mrg { 905 1.1 mrg return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, 906 1.1 mrg (__v64qi) __B, 907 1.1 mrg (__v64qi) __W, 908 1.1 mrg (__mmask64) __U); 909 1.1 mrg } 910 1.1 mrg 911 1.1 mrg extern __inline __m512i 912 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 913 1.1 mrg _mm512_maskz_avg_epu8 (__mmask64 __U, __m512i __A, __m512i __B) 914 1.1 mrg { 915 1.1 mrg return (__m512i) __builtin_ia32_pavgb512_mask ((__v64qi) __A, 916 1.1 mrg (__v64qi) __B, 917 1.1 mrg (__v64qi) 918 1.3 mrg _mm512_setzero_si512 (), 919 1.1 mrg (__mmask64) __U); 920 1.1 mrg } 921 1.1 mrg 922 1.1 mrg extern __inline __m512i 923 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 924 1.1 mrg _mm512_add_epi8 (__m512i __A, __m512i __B) 925 1.1 mrg { 926 1.1 mrg return (__m512i) ((__v64qu) __A + (__v64qu) __B); 927 1.1 mrg } 928 1.1 mrg 929 1.1 mrg extern __inline __m512i 930 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 931 1.1 mrg _mm512_mask_add_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 932 1.1 mrg __m512i __B) 933 1.1 mrg { 934 1.1 mrg return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A, 935 1.1 mrg (__v64qi) __B, 936 1.1 mrg (__v64qi) __W, 937 1.1 mrg (__mmask64) __U); 938 1.1 mrg } 939 1.1 mrg 940 1.1 mrg extern __inline __m512i 941 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 942 1.1 mrg _mm512_maskz_add_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 943 1.1 mrg { 944 1.1 mrg return (__m512i) __builtin_ia32_paddb512_mask ((__v64qi) __A, 945 1.1 mrg (__v64qi) __B, 946 1.1 mrg (__v64qi) 947 1.3 mrg _mm512_setzero_si512 (), 948 1.1 mrg (__mmask64) __U); 949 1.1 mrg } 950 1.1 mrg 951 1.1 mrg extern __inline __m512i 952 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 953 1.1 mrg _mm512_sub_epi8 (__m512i __A, __m512i __B) 954 1.1 mrg { 955 1.1 mrg return (__m512i) ((__v64qu) __A - (__v64qu) __B); 956 1.1 mrg } 957 1.1 mrg 958 1.1 mrg extern __inline __m512i 959 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 960 1.1 mrg _mm512_mask_sub_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 961 1.1 mrg __m512i __B) 962 1.1 mrg { 963 1.1 mrg return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A, 964 1.1 mrg (__v64qi) __B, 965 1.1 mrg (__v64qi) __W, 966 1.1 mrg (__mmask64) __U); 967 1.1 mrg } 968 1.1 mrg 969 1.1 mrg extern __inline __m512i 970 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 971 1.1 mrg _mm512_maskz_sub_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 972 1.1 mrg { 973 1.1 mrg return (__m512i) __builtin_ia32_psubb512_mask ((__v64qi) __A, 974 1.1 mrg (__v64qi) __B, 975 1.1 mrg (__v64qi) 976 1.3 mrg _mm512_setzero_si512 (), 977 1.1 mrg (__mmask64) __U); 978 1.1 mrg } 979 1.1 mrg 980 1.1 mrg extern __inline __m512i 981 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 982 1.1 mrg _mm512_avg_epu16 (__m512i __A, __m512i __B) 983 1.1 mrg { 984 1.1 mrg return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, 985 1.1 mrg (__v32hi) __B, 986 1.1 mrg (__v32hi) 987 1.3 mrg _mm512_setzero_si512 (), 988 1.1 mrg (__mmask32) -1); 989 1.1 mrg } 990 1.1 mrg 991 1.1 mrg extern __inline __m512i 992 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 993 1.1 mrg _mm512_mask_avg_epu16 (__m512i __W, __mmask32 __U, __m512i __A, 994 1.1 mrg __m512i __B) 995 1.1 mrg { 996 1.1 mrg return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, 997 1.1 mrg (__v32hi) __B, 998 1.1 mrg (__v32hi) __W, 999 1.1 mrg (__mmask32) __U); 1000 1.1 mrg } 1001 1.1 mrg 1002 1.1 mrg extern __inline __m512i 1003 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1004 1.1 mrg _mm512_maskz_avg_epu16 (__mmask32 __U, __m512i __A, __m512i __B) 1005 1.1 mrg { 1006 1.1 mrg return (__m512i) __builtin_ia32_pavgw512_mask ((__v32hi) __A, 1007 1.1 mrg (__v32hi) __B, 1008 1.1 mrg (__v32hi) 1009 1.3 mrg _mm512_setzero_si512 (), 1010 1.1 mrg (__mmask32) __U); 1011 1.1 mrg } 1012 1.1 mrg 1013 1.1 mrg extern __inline __m512i 1014 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1015 1.1 mrg _mm512_subs_epi8 (__m512i __A, __m512i __B) 1016 1.1 mrg { 1017 1.1 mrg return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, 1018 1.1 mrg (__v64qi) __B, 1019 1.1 mrg (__v64qi) 1020 1.3 mrg _mm512_setzero_si512 (), 1021 1.1 mrg (__mmask64) -1); 1022 1.1 mrg } 1023 1.1 mrg 1024 1.1 mrg extern __inline __m512i 1025 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1026 1.1 mrg _mm512_mask_subs_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 1027 1.1 mrg __m512i __B) 1028 1.1 mrg { 1029 1.1 mrg return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, 1030 1.1 mrg (__v64qi) __B, 1031 1.1 mrg (__v64qi) __W, 1032 1.1 mrg (__mmask64) __U); 1033 1.1 mrg } 1034 1.1 mrg 1035 1.1 mrg extern __inline __m512i 1036 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1037 1.1 mrg _mm512_maskz_subs_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 1038 1.1 mrg { 1039 1.1 mrg return (__m512i) __builtin_ia32_psubsb512_mask ((__v64qi) __A, 1040 1.1 mrg (__v64qi) __B, 1041 1.1 mrg (__v64qi) 1042 1.3 mrg _mm512_setzero_si512 (), 1043 1.1 mrg (__mmask64) __U); 1044 1.1 mrg } 1045 1.1 mrg 1046 1.1 mrg extern __inline __m512i 1047 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1048 1.1 mrg _mm512_subs_epu8 (__m512i __A, __m512i __B) 1049 1.1 mrg { 1050 1.1 mrg return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, 1051 1.1 mrg (__v64qi) __B, 1052 1.1 mrg (__v64qi) 1053 1.3 mrg _mm512_setzero_si512 (), 1054 1.1 mrg (__mmask64) -1); 1055 1.1 mrg } 1056 1.1 mrg 1057 1.1 mrg extern __inline __m512i 1058 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1059 1.1 mrg _mm512_mask_subs_epu8 (__m512i __W, __mmask64 __U, __m512i __A, 1060 1.1 mrg __m512i __B) 1061 1.1 mrg { 1062 1.1 mrg return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, 1063 1.1 mrg (__v64qi) __B, 1064 1.1 mrg (__v64qi) __W, 1065 1.1 mrg (__mmask64) __U); 1066 1.1 mrg } 1067 1.1 mrg 1068 1.1 mrg extern __inline __m512i 1069 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1070 1.1 mrg _mm512_maskz_subs_epu8 (__mmask64 __U, __m512i __A, __m512i __B) 1071 1.1 mrg { 1072 1.1 mrg return (__m512i) __builtin_ia32_psubusb512_mask ((__v64qi) __A, 1073 1.1 mrg (__v64qi) __B, 1074 1.1 mrg (__v64qi) 1075 1.3 mrg _mm512_setzero_si512 (), 1076 1.1 mrg (__mmask64) __U); 1077 1.1 mrg } 1078 1.1 mrg 1079 1.1 mrg extern __inline __m512i 1080 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1081 1.1 mrg _mm512_adds_epi8 (__m512i __A, __m512i __B) 1082 1.1 mrg { 1083 1.1 mrg return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, 1084 1.1 mrg (__v64qi) __B, 1085 1.1 mrg (__v64qi) 1086 1.3 mrg _mm512_setzero_si512 (), 1087 1.1 mrg (__mmask64) -1); 1088 1.1 mrg } 1089 1.1 mrg 1090 1.1 mrg extern __inline __m512i 1091 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1092 1.1 mrg _mm512_mask_adds_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 1093 1.1 mrg __m512i __B) 1094 1.1 mrg { 1095 1.1 mrg return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, 1096 1.1 mrg (__v64qi) __B, 1097 1.1 mrg (__v64qi) __W, 1098 1.1 mrg (__mmask64) __U); 1099 1.1 mrg } 1100 1.1 mrg 1101 1.1 mrg extern __inline __m512i 1102 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1103 1.1 mrg _mm512_maskz_adds_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 1104 1.1 mrg { 1105 1.1 mrg return (__m512i) __builtin_ia32_paddsb512_mask ((__v64qi) __A, 1106 1.1 mrg (__v64qi) __B, 1107 1.1 mrg (__v64qi) 1108 1.3 mrg _mm512_setzero_si512 (), 1109 1.1 mrg (__mmask64) __U); 1110 1.1 mrg } 1111 1.1 mrg 1112 1.1 mrg extern __inline __m512i 1113 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1114 1.1 mrg _mm512_adds_epu8 (__m512i __A, __m512i __B) 1115 1.1 mrg { 1116 1.1 mrg return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, 1117 1.1 mrg (__v64qi) __B, 1118 1.1 mrg (__v64qi) 1119 1.3 mrg _mm512_setzero_si512 (), 1120 1.1 mrg (__mmask64) -1); 1121 1.1 mrg } 1122 1.1 mrg 1123 1.1 mrg extern __inline __m512i 1124 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1125 1.1 mrg _mm512_mask_adds_epu8 (__m512i __W, __mmask64 __U, __m512i __A, 1126 1.1 mrg __m512i __B) 1127 1.1 mrg { 1128 1.1 mrg return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, 1129 1.1 mrg (__v64qi) __B, 1130 1.1 mrg (__v64qi) __W, 1131 1.1 mrg (__mmask64) __U); 1132 1.1 mrg } 1133 1.1 mrg 1134 1.1 mrg extern __inline __m512i 1135 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1136 1.1 mrg _mm512_maskz_adds_epu8 (__mmask64 __U, __m512i __A, __m512i __B) 1137 1.1 mrg { 1138 1.1 mrg return (__m512i) __builtin_ia32_paddusb512_mask ((__v64qi) __A, 1139 1.1 mrg (__v64qi) __B, 1140 1.1 mrg (__v64qi) 1141 1.3 mrg _mm512_setzero_si512 (), 1142 1.1 mrg (__mmask64) __U); 1143 1.1 mrg } 1144 1.1 mrg 1145 1.1 mrg extern __inline __m512i 1146 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1147 1.1 mrg _mm512_sub_epi16 (__m512i __A, __m512i __B) 1148 1.1 mrg { 1149 1.1 mrg return (__m512i) ((__v32hu) __A - (__v32hu) __B); 1150 1.1 mrg } 1151 1.1 mrg 1152 1.1 mrg extern __inline __m512i 1153 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1154 1.1 mrg _mm512_mask_sub_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1155 1.1 mrg __m512i __B) 1156 1.1 mrg { 1157 1.1 mrg return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A, 1158 1.1 mrg (__v32hi) __B, 1159 1.1 mrg (__v32hi) __W, 1160 1.1 mrg (__mmask32) __U); 1161 1.1 mrg } 1162 1.1 mrg 1163 1.1 mrg extern __inline __m512i 1164 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1165 1.1 mrg _mm512_maskz_sub_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 1166 1.1 mrg { 1167 1.1 mrg return (__m512i) __builtin_ia32_psubw512_mask ((__v32hi) __A, 1168 1.1 mrg (__v32hi) __B, 1169 1.1 mrg (__v32hi) 1170 1.3 mrg _mm512_setzero_si512 (), 1171 1.1 mrg (__mmask32) __U); 1172 1.1 mrg } 1173 1.1 mrg 1174 1.1 mrg extern __inline __m512i 1175 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1176 1.1 mrg _mm512_subs_epi16 (__m512i __A, __m512i __B) 1177 1.1 mrg { 1178 1.1 mrg return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, 1179 1.1 mrg (__v32hi) __B, 1180 1.1 mrg (__v32hi) 1181 1.3 mrg _mm512_setzero_si512 (), 1182 1.1 mrg (__mmask32) -1); 1183 1.1 mrg } 1184 1.1 mrg 1185 1.1 mrg extern __inline __m512i 1186 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1187 1.1 mrg _mm512_mask_subs_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1188 1.1 mrg __m512i __B) 1189 1.1 mrg { 1190 1.1 mrg return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, 1191 1.1 mrg (__v32hi) __B, 1192 1.1 mrg (__v32hi) __W, 1193 1.1 mrg (__mmask32) __U); 1194 1.1 mrg } 1195 1.1 mrg 1196 1.1 mrg extern __inline __m512i 1197 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1198 1.1 mrg _mm512_maskz_subs_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 1199 1.1 mrg { 1200 1.1 mrg return (__m512i) __builtin_ia32_psubsw512_mask ((__v32hi) __A, 1201 1.1 mrg (__v32hi) __B, 1202 1.1 mrg (__v32hi) 1203 1.3 mrg _mm512_setzero_si512 (), 1204 1.1 mrg (__mmask32) __U); 1205 1.1 mrg } 1206 1.1 mrg 1207 1.1 mrg extern __inline __m512i 1208 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1209 1.1 mrg _mm512_subs_epu16 (__m512i __A, __m512i __B) 1210 1.1 mrg { 1211 1.1 mrg return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, 1212 1.1 mrg (__v32hi) __B, 1213 1.1 mrg (__v32hi) 1214 1.3 mrg _mm512_setzero_si512 (), 1215 1.1 mrg (__mmask32) -1); 1216 1.1 mrg } 1217 1.1 mrg 1218 1.1 mrg extern __inline __m512i 1219 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1220 1.1 mrg _mm512_mask_subs_epu16 (__m512i __W, __mmask32 __U, __m512i __A, 1221 1.1 mrg __m512i __B) 1222 1.1 mrg { 1223 1.1 mrg return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, 1224 1.1 mrg (__v32hi) __B, 1225 1.1 mrg (__v32hi) __W, 1226 1.1 mrg (__mmask32) __U); 1227 1.1 mrg } 1228 1.1 mrg 1229 1.1 mrg extern __inline __m512i 1230 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1231 1.1 mrg _mm512_maskz_subs_epu16 (__mmask32 __U, __m512i __A, __m512i __B) 1232 1.1 mrg { 1233 1.1 mrg return (__m512i) __builtin_ia32_psubusw512_mask ((__v32hi) __A, 1234 1.1 mrg (__v32hi) __B, 1235 1.1 mrg (__v32hi) 1236 1.3 mrg _mm512_setzero_si512 (), 1237 1.1 mrg (__mmask32) __U); 1238 1.1 mrg } 1239 1.1 mrg 1240 1.1 mrg extern __inline __m512i 1241 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1242 1.1 mrg _mm512_add_epi16 (__m512i __A, __m512i __B) 1243 1.1 mrg { 1244 1.1 mrg return (__m512i) ((__v32hu) __A + (__v32hu) __B); 1245 1.1 mrg } 1246 1.1 mrg 1247 1.1 mrg extern __inline __m512i 1248 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1249 1.1 mrg _mm512_mask_add_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1250 1.1 mrg __m512i __B) 1251 1.1 mrg { 1252 1.1 mrg return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A, 1253 1.1 mrg (__v32hi) __B, 1254 1.1 mrg (__v32hi) __W, 1255 1.1 mrg (__mmask32) __U); 1256 1.1 mrg } 1257 1.1 mrg 1258 1.1 mrg extern __inline __m512i 1259 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1260 1.1 mrg _mm512_maskz_add_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 1261 1.1 mrg { 1262 1.1 mrg return (__m512i) __builtin_ia32_paddw512_mask ((__v32hi) __A, 1263 1.1 mrg (__v32hi) __B, 1264 1.1 mrg (__v32hi) 1265 1.3 mrg _mm512_setzero_si512 (), 1266 1.1 mrg (__mmask32) __U); 1267 1.1 mrg } 1268 1.1 mrg 1269 1.1 mrg extern __inline __m512i 1270 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1271 1.1 mrg _mm512_adds_epi16 (__m512i __A, __m512i __B) 1272 1.1 mrg { 1273 1.1 mrg return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, 1274 1.1 mrg (__v32hi) __B, 1275 1.1 mrg (__v32hi) 1276 1.3 mrg _mm512_setzero_si512 (), 1277 1.1 mrg (__mmask32) -1); 1278 1.1 mrg } 1279 1.1 mrg 1280 1.1 mrg extern __inline __m512i 1281 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1282 1.1 mrg _mm512_mask_adds_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1283 1.1 mrg __m512i __B) 1284 1.1 mrg { 1285 1.1 mrg return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, 1286 1.1 mrg (__v32hi) __B, 1287 1.1 mrg (__v32hi) __W, 1288 1.1 mrg (__mmask32) __U); 1289 1.1 mrg } 1290 1.1 mrg 1291 1.1 mrg extern __inline __m512i 1292 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1293 1.1 mrg _mm512_maskz_adds_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 1294 1.1 mrg { 1295 1.1 mrg return (__m512i) __builtin_ia32_paddsw512_mask ((__v32hi) __A, 1296 1.1 mrg (__v32hi) __B, 1297 1.1 mrg (__v32hi) 1298 1.3 mrg _mm512_setzero_si512 (), 1299 1.1 mrg (__mmask32) __U); 1300 1.1 mrg } 1301 1.1 mrg 1302 1.1 mrg extern __inline __m512i 1303 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1304 1.1 mrg _mm512_adds_epu16 (__m512i __A, __m512i __B) 1305 1.1 mrg { 1306 1.1 mrg return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, 1307 1.1 mrg (__v32hi) __B, 1308 1.1 mrg (__v32hi) 1309 1.3 mrg _mm512_setzero_si512 (), 1310 1.1 mrg (__mmask32) -1); 1311 1.1 mrg } 1312 1.1 mrg 1313 1.1 mrg extern __inline __m512i 1314 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1315 1.1 mrg _mm512_mask_adds_epu16 (__m512i __W, __mmask32 __U, __m512i __A, 1316 1.1 mrg __m512i __B) 1317 1.1 mrg { 1318 1.1 mrg return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, 1319 1.1 mrg (__v32hi) __B, 1320 1.1 mrg (__v32hi) __W, 1321 1.1 mrg (__mmask32) __U); 1322 1.1 mrg } 1323 1.1 mrg 1324 1.1 mrg extern __inline __m512i 1325 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1326 1.1 mrg _mm512_maskz_adds_epu16 (__mmask32 __U, __m512i __A, __m512i __B) 1327 1.1 mrg { 1328 1.1 mrg return (__m512i) __builtin_ia32_paddusw512_mask ((__v32hi) __A, 1329 1.1 mrg (__v32hi) __B, 1330 1.1 mrg (__v32hi) 1331 1.3 mrg _mm512_setzero_si512 (), 1332 1.1 mrg (__mmask32) __U); 1333 1.1 mrg } 1334 1.1 mrg 1335 1.1 mrg extern __inline __m512i 1336 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1337 1.1 mrg _mm512_srl_epi16 (__m512i __A, __m128i __B) 1338 1.1 mrg { 1339 1.1 mrg return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, 1340 1.1 mrg (__v8hi) __B, 1341 1.1 mrg (__v32hi) 1342 1.3 mrg _mm512_setzero_si512 (), 1343 1.1 mrg (__mmask32) -1); 1344 1.1 mrg } 1345 1.1 mrg 1346 1.1 mrg extern __inline __m512i 1347 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1348 1.1 mrg _mm512_mask_srl_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1349 1.1 mrg __m128i __B) 1350 1.1 mrg { 1351 1.1 mrg return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, 1352 1.1 mrg (__v8hi) __B, 1353 1.1 mrg (__v32hi) __W, 1354 1.1 mrg (__mmask32) __U); 1355 1.1 mrg } 1356 1.1 mrg 1357 1.1 mrg extern __inline __m512i 1358 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1359 1.1 mrg _mm512_maskz_srl_epi16 (__mmask32 __U, __m512i __A, __m128i __B) 1360 1.1 mrg { 1361 1.1 mrg return (__m512i) __builtin_ia32_psrlw512_mask ((__v32hi) __A, 1362 1.1 mrg (__v8hi) __B, 1363 1.1 mrg (__v32hi) 1364 1.3 mrg _mm512_setzero_si512 (), 1365 1.1 mrg (__mmask32) __U); 1366 1.1 mrg } 1367 1.1 mrg 1368 1.1 mrg extern __inline __m512i 1369 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1370 1.1 mrg _mm512_packs_epi16 (__m512i __A, __m512i __B) 1371 1.1 mrg { 1372 1.1 mrg return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, 1373 1.1 mrg (__v32hi) __B, 1374 1.1 mrg (__v64qi) 1375 1.3 mrg _mm512_setzero_si512 (), 1376 1.1 mrg (__mmask64) -1); 1377 1.1 mrg } 1378 1.1 mrg 1379 1.1 mrg extern __inline __m512i 1380 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1381 1.1 mrg _mm512_sll_epi16 (__m512i __A, __m128i __B) 1382 1.1 mrg { 1383 1.1 mrg return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, 1384 1.1 mrg (__v8hi) __B, 1385 1.1 mrg (__v32hi) 1386 1.3 mrg _mm512_setzero_si512 (), 1387 1.1 mrg (__mmask32) -1); 1388 1.1 mrg } 1389 1.1 mrg 1390 1.1 mrg extern __inline __m512i 1391 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1392 1.1 mrg _mm512_mask_sll_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1393 1.1 mrg __m128i __B) 1394 1.1 mrg { 1395 1.1 mrg return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, 1396 1.1 mrg (__v8hi) __B, 1397 1.1 mrg (__v32hi) __W, 1398 1.1 mrg (__mmask32) __U); 1399 1.1 mrg } 1400 1.1 mrg 1401 1.1 mrg extern __inline __m512i 1402 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1403 1.1 mrg _mm512_maskz_sll_epi16 (__mmask32 __U, __m512i __A, __m128i __B) 1404 1.1 mrg { 1405 1.1 mrg return (__m512i) __builtin_ia32_psllw512_mask ((__v32hi) __A, 1406 1.1 mrg (__v8hi) __B, 1407 1.1 mrg (__v32hi) 1408 1.3 mrg _mm512_setzero_si512 (), 1409 1.1 mrg (__mmask32) __U); 1410 1.1 mrg } 1411 1.1 mrg 1412 1.1 mrg extern __inline __m512i 1413 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1414 1.1 mrg _mm512_maddubs_epi16 (__m512i __X, __m512i __Y) 1415 1.1 mrg { 1416 1.1 mrg return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, 1417 1.1 mrg (__v64qi) __Y, 1418 1.1 mrg (__v32hi) 1419 1.3 mrg _mm512_setzero_si512 (), 1420 1.1 mrg (__mmask32) -1); 1421 1.1 mrg } 1422 1.1 mrg 1423 1.1 mrg extern __inline __m512i 1424 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1425 1.1 mrg _mm512_mask_maddubs_epi16 (__m512i __W, __mmask32 __U, __m512i __X, 1426 1.1 mrg __m512i __Y) 1427 1.1 mrg { 1428 1.1 mrg return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, 1429 1.1 mrg (__v64qi) __Y, 1430 1.1 mrg (__v32hi) __W, 1431 1.1 mrg (__mmask32) __U); 1432 1.1 mrg } 1433 1.1 mrg 1434 1.1 mrg extern __inline __m512i 1435 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1436 1.1 mrg _mm512_maskz_maddubs_epi16 (__mmask32 __U, __m512i __X, __m512i __Y) 1437 1.1 mrg { 1438 1.1 mrg return (__m512i) __builtin_ia32_pmaddubsw512_mask ((__v64qi) __X, 1439 1.1 mrg (__v64qi) __Y, 1440 1.1 mrg (__v32hi) 1441 1.3 mrg _mm512_setzero_si512 (), 1442 1.1 mrg (__mmask32) __U); 1443 1.1 mrg } 1444 1.1 mrg 1445 1.1 mrg extern __inline __m512i 1446 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1447 1.1 mrg _mm512_madd_epi16 (__m512i __A, __m512i __B) 1448 1.1 mrg { 1449 1.1 mrg return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, 1450 1.1 mrg (__v32hi) __B, 1451 1.1 mrg (__v16si) 1452 1.1 mrg _mm512_setzero_si512 (), 1453 1.1 mrg (__mmask16) -1); 1454 1.1 mrg } 1455 1.1 mrg 1456 1.1 mrg extern __inline __m512i 1457 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1458 1.1 mrg _mm512_mask_madd_epi16 (__m512i __W, __mmask16 __U, __m512i __A, 1459 1.1 mrg __m512i __B) 1460 1.1 mrg { 1461 1.1 mrg return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, 1462 1.1 mrg (__v32hi) __B, 1463 1.1 mrg (__v16si) __W, 1464 1.1 mrg (__mmask16) __U); 1465 1.1 mrg } 1466 1.1 mrg 1467 1.1 mrg extern __inline __m512i 1468 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1469 1.1 mrg _mm512_maskz_madd_epi16 (__mmask16 __U, __m512i __A, __m512i __B) 1470 1.1 mrg { 1471 1.1 mrg return (__m512i) __builtin_ia32_pmaddwd512_mask ((__v32hi) __A, 1472 1.1 mrg (__v32hi) __B, 1473 1.1 mrg (__v16si) 1474 1.1 mrg _mm512_setzero_si512 (), 1475 1.1 mrg (__mmask16) __U); 1476 1.1 mrg } 1477 1.1 mrg 1478 1.1 mrg extern __inline __m512i 1479 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1480 1.1 mrg _mm512_unpackhi_epi8 (__m512i __A, __m512i __B) 1481 1.1 mrg { 1482 1.1 mrg return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, 1483 1.1 mrg (__v64qi) __B, 1484 1.1 mrg (__v64qi) 1485 1.3 mrg _mm512_setzero_si512 (), 1486 1.1 mrg (__mmask64) -1); 1487 1.1 mrg } 1488 1.1 mrg 1489 1.1 mrg extern __inline __m512i 1490 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1491 1.1 mrg _mm512_mask_unpackhi_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 1492 1.1 mrg __m512i __B) 1493 1.1 mrg { 1494 1.1 mrg return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, 1495 1.1 mrg (__v64qi) __B, 1496 1.1 mrg (__v64qi) __W, 1497 1.1 mrg (__mmask64) __U); 1498 1.1 mrg } 1499 1.1 mrg 1500 1.1 mrg extern __inline __m512i 1501 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1502 1.1 mrg _mm512_maskz_unpackhi_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 1503 1.1 mrg { 1504 1.1 mrg return (__m512i) __builtin_ia32_punpckhbw512_mask ((__v64qi) __A, 1505 1.1 mrg (__v64qi) __B, 1506 1.1 mrg (__v64qi) 1507 1.3 mrg _mm512_setzero_si512 (), 1508 1.1 mrg (__mmask64) __U); 1509 1.1 mrg } 1510 1.1 mrg 1511 1.1 mrg extern __inline __m512i 1512 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1513 1.1 mrg _mm512_unpackhi_epi16 (__m512i __A, __m512i __B) 1514 1.1 mrg { 1515 1.1 mrg return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, 1516 1.1 mrg (__v32hi) __B, 1517 1.1 mrg (__v32hi) 1518 1.3 mrg _mm512_setzero_si512 (), 1519 1.1 mrg (__mmask32) -1); 1520 1.1 mrg } 1521 1.1 mrg 1522 1.1 mrg extern __inline __m512i 1523 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1524 1.1 mrg _mm512_mask_unpackhi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1525 1.1 mrg __m512i __B) 1526 1.1 mrg { 1527 1.1 mrg return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, 1528 1.1 mrg (__v32hi) __B, 1529 1.1 mrg (__v32hi) __W, 1530 1.1 mrg (__mmask32) __U); 1531 1.1 mrg } 1532 1.1 mrg 1533 1.1 mrg extern __inline __m512i 1534 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1535 1.1 mrg _mm512_maskz_unpackhi_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 1536 1.1 mrg { 1537 1.1 mrg return (__m512i) __builtin_ia32_punpckhwd512_mask ((__v32hi) __A, 1538 1.1 mrg (__v32hi) __B, 1539 1.1 mrg (__v32hi) 1540 1.3 mrg _mm512_setzero_si512 (), 1541 1.1 mrg (__mmask32) __U); 1542 1.1 mrg } 1543 1.1 mrg 1544 1.1 mrg extern __inline __m512i 1545 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1546 1.1 mrg _mm512_unpacklo_epi8 (__m512i __A, __m512i __B) 1547 1.1 mrg { 1548 1.1 mrg return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, 1549 1.1 mrg (__v64qi) __B, 1550 1.1 mrg (__v64qi) 1551 1.3 mrg _mm512_setzero_si512 (), 1552 1.1 mrg (__mmask64) -1); 1553 1.1 mrg } 1554 1.1 mrg 1555 1.1 mrg extern __inline __m512i 1556 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1557 1.1 mrg _mm512_mask_unpacklo_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 1558 1.1 mrg __m512i __B) 1559 1.1 mrg { 1560 1.1 mrg return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, 1561 1.1 mrg (__v64qi) __B, 1562 1.1 mrg (__v64qi) __W, 1563 1.1 mrg (__mmask64) __U); 1564 1.1 mrg } 1565 1.1 mrg 1566 1.1 mrg extern __inline __m512i 1567 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1568 1.1 mrg _mm512_maskz_unpacklo_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 1569 1.1 mrg { 1570 1.1 mrg return (__m512i) __builtin_ia32_punpcklbw512_mask ((__v64qi) __A, 1571 1.1 mrg (__v64qi) __B, 1572 1.1 mrg (__v64qi) 1573 1.3 mrg _mm512_setzero_si512 (), 1574 1.1 mrg (__mmask64) __U); 1575 1.1 mrg } 1576 1.1 mrg 1577 1.1 mrg extern __inline __m512i 1578 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1579 1.1 mrg _mm512_unpacklo_epi16 (__m512i __A, __m512i __B) 1580 1.1 mrg { 1581 1.1 mrg return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, 1582 1.1 mrg (__v32hi) __B, 1583 1.1 mrg (__v32hi) 1584 1.3 mrg _mm512_setzero_si512 (), 1585 1.1 mrg (__mmask32) -1); 1586 1.1 mrg } 1587 1.1 mrg 1588 1.1 mrg extern __inline __m512i 1589 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1590 1.1 mrg _mm512_mask_unpacklo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 1591 1.1 mrg __m512i __B) 1592 1.1 mrg { 1593 1.1 mrg return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, 1594 1.1 mrg (__v32hi) __B, 1595 1.1 mrg (__v32hi) __W, 1596 1.1 mrg (__mmask32) __U); 1597 1.1 mrg } 1598 1.1 mrg 1599 1.1 mrg extern __inline __m512i 1600 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1601 1.1 mrg _mm512_maskz_unpacklo_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 1602 1.1 mrg { 1603 1.1 mrg return (__m512i) __builtin_ia32_punpcklwd512_mask ((__v32hi) __A, 1604 1.1 mrg (__v32hi) __B, 1605 1.1 mrg (__v32hi) 1606 1.3 mrg _mm512_setzero_si512 (), 1607 1.1 mrg (__mmask32) __U); 1608 1.1 mrg } 1609 1.1 mrg 1610 1.1 mrg extern __inline __mmask64 1611 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1612 1.1 mrg _mm512_cmpeq_epu8_mask (__m512i __A, __m512i __B) 1613 1.1 mrg { 1614 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, 1615 1.1 mrg (__v64qi) __B, 0, 1616 1.1 mrg (__mmask64) -1); 1617 1.1 mrg } 1618 1.1 mrg 1619 1.1 mrg extern __inline __mmask64 1620 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1621 1.1 mrg _mm512_cmpeq_epi8_mask (__m512i __A, __m512i __B) 1622 1.1 mrg { 1623 1.1 mrg return (__mmask64) __builtin_ia32_pcmpeqb512_mask ((__v64qi) __A, 1624 1.1 mrg (__v64qi) __B, 1625 1.1 mrg (__mmask64) -1); 1626 1.1 mrg } 1627 1.1 mrg 1628 1.1 mrg extern __inline __mmask64 1629 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1630 1.1 mrg _mm512_mask_cmpeq_epu8_mask (__mmask64 __U, __m512i __A, __m512i __B) 1631 1.1 mrg { 1632 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, 1633 1.1 mrg (__v64qi) __B, 0, 1634 1.1 mrg __U); 1635 1.1 mrg } 1636 1.1 mrg 1637 1.1 mrg extern __inline __mmask64 1638 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1639 1.1 mrg _mm512_mask_cmpeq_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) 1640 1.1 mrg { 1641 1.1 mrg return (__mmask64) __builtin_ia32_pcmpeqb512_mask ((__v64qi) __A, 1642 1.1 mrg (__v64qi) __B, 1643 1.1 mrg __U); 1644 1.1 mrg } 1645 1.1 mrg 1646 1.1 mrg extern __inline __mmask32 1647 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1648 1.1 mrg _mm512_cmpeq_epu16_mask (__m512i __A, __m512i __B) 1649 1.1 mrg { 1650 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, 1651 1.1 mrg (__v32hi) __B, 0, 1652 1.1 mrg (__mmask32) -1); 1653 1.1 mrg } 1654 1.1 mrg 1655 1.1 mrg extern __inline __mmask32 1656 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1657 1.1 mrg _mm512_cmpeq_epi16_mask (__m512i __A, __m512i __B) 1658 1.1 mrg { 1659 1.1 mrg return (__mmask32) __builtin_ia32_pcmpeqw512_mask ((__v32hi) __A, 1660 1.1 mrg (__v32hi) __B, 1661 1.1 mrg (__mmask32) -1); 1662 1.1 mrg } 1663 1.1 mrg 1664 1.1 mrg extern __inline __mmask32 1665 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1666 1.1 mrg _mm512_mask_cmpeq_epu16_mask (__mmask32 __U, __m512i __A, __m512i __B) 1667 1.1 mrg { 1668 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, 1669 1.1 mrg (__v32hi) __B, 0, 1670 1.1 mrg __U); 1671 1.1 mrg } 1672 1.1 mrg 1673 1.1 mrg extern __inline __mmask32 1674 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1675 1.1 mrg _mm512_mask_cmpeq_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) 1676 1.1 mrg { 1677 1.1 mrg return (__mmask32) __builtin_ia32_pcmpeqw512_mask ((__v32hi) __A, 1678 1.1 mrg (__v32hi) __B, 1679 1.1 mrg __U); 1680 1.1 mrg } 1681 1.1 mrg 1682 1.1 mrg extern __inline __mmask64 1683 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1684 1.1 mrg _mm512_cmpgt_epu8_mask (__m512i __A, __m512i __B) 1685 1.1 mrg { 1686 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, 1687 1.1 mrg (__v64qi) __B, 6, 1688 1.1 mrg (__mmask64) -1); 1689 1.1 mrg } 1690 1.1 mrg 1691 1.1 mrg extern __inline __mmask64 1692 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1693 1.1 mrg _mm512_cmpgt_epi8_mask (__m512i __A, __m512i __B) 1694 1.1 mrg { 1695 1.1 mrg return (__mmask64) __builtin_ia32_pcmpgtb512_mask ((__v64qi) __A, 1696 1.1 mrg (__v64qi) __B, 1697 1.1 mrg (__mmask64) -1); 1698 1.1 mrg } 1699 1.1 mrg 1700 1.1 mrg extern __inline __mmask64 1701 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1702 1.1 mrg _mm512_mask_cmpgt_epu8_mask (__mmask64 __U, __m512i __A, __m512i __B) 1703 1.1 mrg { 1704 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __A, 1705 1.1 mrg (__v64qi) __B, 6, 1706 1.1 mrg __U); 1707 1.1 mrg } 1708 1.1 mrg 1709 1.1 mrg extern __inline __mmask64 1710 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1711 1.1 mrg _mm512_mask_cmpgt_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) 1712 1.1 mrg { 1713 1.1 mrg return (__mmask64) __builtin_ia32_pcmpgtb512_mask ((__v64qi) __A, 1714 1.1 mrg (__v64qi) __B, 1715 1.1 mrg __U); 1716 1.1 mrg } 1717 1.1 mrg 1718 1.1 mrg extern __inline __mmask32 1719 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1720 1.1 mrg _mm512_cmpgt_epu16_mask (__m512i __A, __m512i __B) 1721 1.1 mrg { 1722 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, 1723 1.1 mrg (__v32hi) __B, 6, 1724 1.1 mrg (__mmask32) -1); 1725 1.1 mrg } 1726 1.1 mrg 1727 1.1 mrg extern __inline __mmask32 1728 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1729 1.1 mrg _mm512_cmpgt_epi16_mask (__m512i __A, __m512i __B) 1730 1.1 mrg { 1731 1.1 mrg return (__mmask32) __builtin_ia32_pcmpgtw512_mask ((__v32hi) __A, 1732 1.1 mrg (__v32hi) __B, 1733 1.1 mrg (__mmask32) -1); 1734 1.1 mrg } 1735 1.1 mrg 1736 1.1 mrg extern __inline __mmask32 1737 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1738 1.1 mrg _mm512_mask_cmpgt_epu16_mask (__mmask32 __U, __m512i __A, __m512i __B) 1739 1.1 mrg { 1740 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __A, 1741 1.1 mrg (__v32hi) __B, 6, 1742 1.1 mrg __U); 1743 1.1 mrg } 1744 1.1 mrg 1745 1.1 mrg extern __inline __mmask32 1746 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1747 1.1 mrg _mm512_mask_cmpgt_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) 1748 1.1 mrg { 1749 1.1 mrg return (__mmask32) __builtin_ia32_pcmpgtw512_mask ((__v32hi) __A, 1750 1.1 mrg (__v32hi) __B, 1751 1.1 mrg __U); 1752 1.1 mrg } 1753 1.1 mrg 1754 1.1 mrg extern __inline __mmask64 1755 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1756 1.1 mrg _mm512_movepi8_mask (__m512i __A) 1757 1.1 mrg { 1758 1.1 mrg return (__mmask64) __builtin_ia32_cvtb2mask512 ((__v64qi) __A); 1759 1.1 mrg } 1760 1.1 mrg 1761 1.1 mrg extern __inline __mmask32 1762 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1763 1.1 mrg _mm512_movepi16_mask (__m512i __A) 1764 1.1 mrg { 1765 1.1 mrg return (__mmask32) __builtin_ia32_cvtw2mask512 ((__v32hi) __A); 1766 1.1 mrg } 1767 1.1 mrg 1768 1.1 mrg extern __inline __m512i 1769 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1770 1.1 mrg _mm512_movm_epi8 (__mmask64 __A) 1771 1.1 mrg { 1772 1.1 mrg return (__m512i) __builtin_ia32_cvtmask2b512 (__A); 1773 1.1 mrg } 1774 1.1 mrg 1775 1.1 mrg extern __inline __m512i 1776 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1777 1.1 mrg _mm512_movm_epi16 (__mmask32 __A) 1778 1.1 mrg { 1779 1.1 mrg return (__m512i) __builtin_ia32_cvtmask2w512 (__A); 1780 1.1 mrg } 1781 1.1 mrg 1782 1.1 mrg extern __inline __mmask64 1783 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1784 1.1 mrg _mm512_test_epi8_mask (__m512i __A, __m512i __B) 1785 1.1 mrg { 1786 1.1 mrg return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A, 1787 1.1 mrg (__v64qi) __B, 1788 1.1 mrg (__mmask64) -1); 1789 1.1 mrg } 1790 1.1 mrg 1791 1.1 mrg extern __inline __mmask64 1792 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1793 1.1 mrg _mm512_mask_test_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) 1794 1.1 mrg { 1795 1.1 mrg return (__mmask64) __builtin_ia32_ptestmb512 ((__v64qi) __A, 1796 1.1 mrg (__v64qi) __B, __U); 1797 1.1 mrg } 1798 1.1 mrg 1799 1.1 mrg extern __inline __mmask32 1800 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1801 1.1 mrg _mm512_test_epi16_mask (__m512i __A, __m512i __B) 1802 1.1 mrg { 1803 1.1 mrg return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A, 1804 1.1 mrg (__v32hi) __B, 1805 1.1 mrg (__mmask32) -1); 1806 1.1 mrg } 1807 1.1 mrg 1808 1.1 mrg extern __inline __mmask32 1809 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1810 1.1 mrg _mm512_mask_test_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) 1811 1.1 mrg { 1812 1.1 mrg return (__mmask32) __builtin_ia32_ptestmw512 ((__v32hi) __A, 1813 1.1 mrg (__v32hi) __B, __U); 1814 1.1 mrg } 1815 1.1 mrg 1816 1.1 mrg extern __inline __mmask64 1817 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1818 1.1 mrg _mm512_testn_epi8_mask (__m512i __A, __m512i __B) 1819 1.1 mrg { 1820 1.1 mrg return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A, 1821 1.1 mrg (__v64qi) __B, 1822 1.1 mrg (__mmask64) -1); 1823 1.1 mrg } 1824 1.1 mrg 1825 1.1 mrg extern __inline __mmask64 1826 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1827 1.1 mrg _mm512_mask_testn_epi8_mask (__mmask64 __U, __m512i __A, __m512i __B) 1828 1.1 mrg { 1829 1.1 mrg return (__mmask64) __builtin_ia32_ptestnmb512 ((__v64qi) __A, 1830 1.1 mrg (__v64qi) __B, __U); 1831 1.1 mrg } 1832 1.1 mrg 1833 1.1 mrg extern __inline __mmask32 1834 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1835 1.1 mrg _mm512_testn_epi16_mask (__m512i __A, __m512i __B) 1836 1.1 mrg { 1837 1.1 mrg return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A, 1838 1.1 mrg (__v32hi) __B, 1839 1.1 mrg (__mmask32) -1); 1840 1.1 mrg } 1841 1.1 mrg 1842 1.1 mrg extern __inline __mmask32 1843 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1844 1.1 mrg _mm512_mask_testn_epi16_mask (__mmask32 __U, __m512i __A, __m512i __B) 1845 1.1 mrg { 1846 1.1 mrg return (__mmask32) __builtin_ia32_ptestnmw512 ((__v32hi) __A, 1847 1.1 mrg (__v32hi) __B, __U); 1848 1.1 mrg } 1849 1.1 mrg 1850 1.1 mrg extern __inline __m512i 1851 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1852 1.1 mrg _mm512_shuffle_epi8 (__m512i __A, __m512i __B) 1853 1.1 mrg { 1854 1.1 mrg return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, 1855 1.1 mrg (__v64qi) __B, 1856 1.1 mrg (__v64qi) 1857 1.3 mrg _mm512_setzero_si512 (), 1858 1.1 mrg (__mmask64) -1); 1859 1.1 mrg } 1860 1.1 mrg 1861 1.1 mrg extern __inline __m512i 1862 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1863 1.1 mrg _mm512_mask_shuffle_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 1864 1.1 mrg __m512i __B) 1865 1.1 mrg { 1866 1.1 mrg return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, 1867 1.1 mrg (__v64qi) __B, 1868 1.1 mrg (__v64qi) __W, 1869 1.1 mrg (__mmask64) __U); 1870 1.1 mrg } 1871 1.1 mrg 1872 1.1 mrg extern __inline __m512i 1873 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1874 1.1 mrg _mm512_maskz_shuffle_epi8 (__mmask64 __U, __m512i __A, __m512i __B) 1875 1.1 mrg { 1876 1.1 mrg return (__m512i) __builtin_ia32_pshufb512_mask ((__v64qi) __A, 1877 1.1 mrg (__v64qi) __B, 1878 1.1 mrg (__v64qi) 1879 1.3 mrg _mm512_setzero_si512 (), 1880 1.1 mrg (__mmask64) __U); 1881 1.1 mrg } 1882 1.1 mrg 1883 1.1 mrg extern __inline __m512i 1884 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1885 1.1 mrg _mm512_min_epu16 (__m512i __A, __m512i __B) 1886 1.1 mrg { 1887 1.1 mrg return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, 1888 1.1 mrg (__v32hi) __B, 1889 1.1 mrg (__v32hi) 1890 1.3 mrg _mm512_setzero_si512 (), 1891 1.1 mrg (__mmask32) -1); 1892 1.1 mrg } 1893 1.1 mrg 1894 1.1 mrg extern __inline __m512i 1895 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1896 1.1 mrg _mm512_maskz_min_epu16 (__mmask32 __M, __m512i __A, __m512i __B) 1897 1.1 mrg { 1898 1.1 mrg return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, 1899 1.1 mrg (__v32hi) __B, 1900 1.1 mrg (__v32hi) 1901 1.3 mrg _mm512_setzero_si512 (), 1902 1.1 mrg (__mmask32) __M); 1903 1.1 mrg } 1904 1.1 mrg 1905 1.1 mrg extern __inline __m512i 1906 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1907 1.1 mrg _mm512_mask_min_epu16 (__m512i __W, __mmask32 __M, __m512i __A, 1908 1.1 mrg __m512i __B) 1909 1.1 mrg { 1910 1.1 mrg return (__m512i) __builtin_ia32_pminuw512_mask ((__v32hi) __A, 1911 1.1 mrg (__v32hi) __B, 1912 1.1 mrg (__v32hi) __W, 1913 1.1 mrg (__mmask32) __M); 1914 1.1 mrg } 1915 1.1 mrg 1916 1.1 mrg extern __inline __m512i 1917 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1918 1.1 mrg _mm512_min_epi16 (__m512i __A, __m512i __B) 1919 1.1 mrg { 1920 1.1 mrg return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, 1921 1.1 mrg (__v32hi) __B, 1922 1.1 mrg (__v32hi) 1923 1.3 mrg _mm512_setzero_si512 (), 1924 1.1 mrg (__mmask32) -1); 1925 1.1 mrg } 1926 1.1 mrg 1927 1.1 mrg extern __inline __m512i 1928 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1929 1.1 mrg _mm512_maskz_min_epi16 (__mmask32 __M, __m512i __A, __m512i __B) 1930 1.1 mrg { 1931 1.1 mrg return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, 1932 1.1 mrg (__v32hi) __B, 1933 1.1 mrg (__v32hi) 1934 1.3 mrg _mm512_setzero_si512 (), 1935 1.1 mrg (__mmask32) __M); 1936 1.1 mrg } 1937 1.1 mrg 1938 1.1 mrg extern __inline __m512i 1939 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1940 1.1 mrg _mm512_mask_min_epi16 (__m512i __W, __mmask32 __M, __m512i __A, 1941 1.1 mrg __m512i __B) 1942 1.1 mrg { 1943 1.1 mrg return (__m512i) __builtin_ia32_pminsw512_mask ((__v32hi) __A, 1944 1.1 mrg (__v32hi) __B, 1945 1.1 mrg (__v32hi) __W, 1946 1.1 mrg (__mmask32) __M); 1947 1.1 mrg } 1948 1.1 mrg 1949 1.1 mrg extern __inline __m512i 1950 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1951 1.1 mrg _mm512_max_epu8 (__m512i __A, __m512i __B) 1952 1.1 mrg { 1953 1.1 mrg return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, 1954 1.1 mrg (__v64qi) __B, 1955 1.1 mrg (__v64qi) 1956 1.3 mrg _mm512_setzero_si512 (), 1957 1.1 mrg (__mmask64) -1); 1958 1.1 mrg } 1959 1.1 mrg 1960 1.1 mrg extern __inline __m512i 1961 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1962 1.1 mrg _mm512_maskz_max_epu8 (__mmask64 __M, __m512i __A, __m512i __B) 1963 1.1 mrg { 1964 1.1 mrg return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, 1965 1.1 mrg (__v64qi) __B, 1966 1.1 mrg (__v64qi) 1967 1.3 mrg _mm512_setzero_si512 (), 1968 1.1 mrg (__mmask64) __M); 1969 1.1 mrg } 1970 1.1 mrg 1971 1.1 mrg extern __inline __m512i 1972 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1973 1.1 mrg _mm512_mask_max_epu8 (__m512i __W, __mmask64 __M, __m512i __A, 1974 1.1 mrg __m512i __B) 1975 1.1 mrg { 1976 1.1 mrg return (__m512i) __builtin_ia32_pmaxub512_mask ((__v64qi) __A, 1977 1.1 mrg (__v64qi) __B, 1978 1.1 mrg (__v64qi) __W, 1979 1.1 mrg (__mmask64) __M); 1980 1.1 mrg } 1981 1.1 mrg 1982 1.1 mrg extern __inline __m512i 1983 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1984 1.1 mrg _mm512_max_epi8 (__m512i __A, __m512i __B) 1985 1.1 mrg { 1986 1.1 mrg return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, 1987 1.1 mrg (__v64qi) __B, 1988 1.1 mrg (__v64qi) 1989 1.3 mrg _mm512_setzero_si512 (), 1990 1.1 mrg (__mmask64) -1); 1991 1.1 mrg } 1992 1.1 mrg 1993 1.1 mrg extern __inline __m512i 1994 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1995 1.1 mrg _mm512_maskz_max_epi8 (__mmask64 __M, __m512i __A, __m512i __B) 1996 1.1 mrg { 1997 1.1 mrg return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, 1998 1.1 mrg (__v64qi) __B, 1999 1.1 mrg (__v64qi) 2000 1.3 mrg _mm512_setzero_si512 (), 2001 1.1 mrg (__mmask64) __M); 2002 1.1 mrg } 2003 1.1 mrg 2004 1.1 mrg extern __inline __m512i 2005 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2006 1.1 mrg _mm512_mask_max_epi8 (__m512i __W, __mmask64 __M, __m512i __A, 2007 1.1 mrg __m512i __B) 2008 1.1 mrg { 2009 1.1 mrg return (__m512i) __builtin_ia32_pmaxsb512_mask ((__v64qi) __A, 2010 1.1 mrg (__v64qi) __B, 2011 1.1 mrg (__v64qi) __W, 2012 1.1 mrg (__mmask64) __M); 2013 1.1 mrg } 2014 1.1 mrg 2015 1.1 mrg extern __inline __m512i 2016 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2017 1.1 mrg _mm512_min_epu8 (__m512i __A, __m512i __B) 2018 1.1 mrg { 2019 1.1 mrg return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, 2020 1.1 mrg (__v64qi) __B, 2021 1.1 mrg (__v64qi) 2022 1.3 mrg _mm512_setzero_si512 (), 2023 1.1 mrg (__mmask64) -1); 2024 1.1 mrg } 2025 1.1 mrg 2026 1.1 mrg extern __inline __m512i 2027 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2028 1.1 mrg _mm512_maskz_min_epu8 (__mmask64 __M, __m512i __A, __m512i __B) 2029 1.1 mrg { 2030 1.1 mrg return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, 2031 1.1 mrg (__v64qi) __B, 2032 1.1 mrg (__v64qi) 2033 1.3 mrg _mm512_setzero_si512 (), 2034 1.1 mrg (__mmask64) __M); 2035 1.1 mrg } 2036 1.1 mrg 2037 1.1 mrg extern __inline __m512i 2038 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2039 1.1 mrg _mm512_mask_min_epu8 (__m512i __W, __mmask64 __M, __m512i __A, 2040 1.1 mrg __m512i __B) 2041 1.1 mrg { 2042 1.1 mrg return (__m512i) __builtin_ia32_pminub512_mask ((__v64qi) __A, 2043 1.1 mrg (__v64qi) __B, 2044 1.1 mrg (__v64qi) __W, 2045 1.1 mrg (__mmask64) __M); 2046 1.1 mrg } 2047 1.1 mrg 2048 1.1 mrg extern __inline __m512i 2049 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2050 1.1 mrg _mm512_min_epi8 (__m512i __A, __m512i __B) 2051 1.1 mrg { 2052 1.1 mrg return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, 2053 1.1 mrg (__v64qi) __B, 2054 1.1 mrg (__v64qi) 2055 1.3 mrg _mm512_setzero_si512 (), 2056 1.1 mrg (__mmask64) -1); 2057 1.1 mrg } 2058 1.1 mrg 2059 1.1 mrg extern __inline __m512i 2060 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2061 1.1 mrg _mm512_maskz_min_epi8 (__mmask64 __M, __m512i __A, __m512i __B) 2062 1.1 mrg { 2063 1.1 mrg return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, 2064 1.1 mrg (__v64qi) __B, 2065 1.1 mrg (__v64qi) 2066 1.3 mrg _mm512_setzero_si512 (), 2067 1.1 mrg (__mmask64) __M); 2068 1.1 mrg } 2069 1.1 mrg 2070 1.1 mrg extern __inline __m512i 2071 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2072 1.1 mrg _mm512_mask_min_epi8 (__m512i __W, __mmask64 __M, __m512i __A, 2073 1.1 mrg __m512i __B) 2074 1.1 mrg { 2075 1.1 mrg return (__m512i) __builtin_ia32_pminsb512_mask ((__v64qi) __A, 2076 1.1 mrg (__v64qi) __B, 2077 1.1 mrg (__v64qi) __W, 2078 1.1 mrg (__mmask64) __M); 2079 1.1 mrg } 2080 1.1 mrg 2081 1.1 mrg extern __inline __m512i 2082 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2083 1.1 mrg _mm512_max_epi16 (__m512i __A, __m512i __B) 2084 1.1 mrg { 2085 1.1 mrg return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, 2086 1.1 mrg (__v32hi) __B, 2087 1.1 mrg (__v32hi) 2088 1.3 mrg _mm512_setzero_si512 (), 2089 1.1 mrg (__mmask32) -1); 2090 1.1 mrg } 2091 1.1 mrg 2092 1.1 mrg extern __inline __m512i 2093 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2094 1.1 mrg _mm512_maskz_max_epi16 (__mmask32 __M, __m512i __A, __m512i __B) 2095 1.1 mrg { 2096 1.1 mrg return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, 2097 1.1 mrg (__v32hi) __B, 2098 1.1 mrg (__v32hi) 2099 1.3 mrg _mm512_setzero_si512 (), 2100 1.1 mrg (__mmask32) __M); 2101 1.1 mrg } 2102 1.1 mrg 2103 1.1 mrg extern __inline __m512i 2104 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2105 1.1 mrg _mm512_mask_max_epi16 (__m512i __W, __mmask32 __M, __m512i __A, 2106 1.1 mrg __m512i __B) 2107 1.1 mrg { 2108 1.1 mrg return (__m512i) __builtin_ia32_pmaxsw512_mask ((__v32hi) __A, 2109 1.1 mrg (__v32hi) __B, 2110 1.1 mrg (__v32hi) __W, 2111 1.1 mrg (__mmask32) __M); 2112 1.1 mrg } 2113 1.1 mrg 2114 1.1 mrg extern __inline __m512i 2115 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2116 1.1 mrg _mm512_max_epu16 (__m512i __A, __m512i __B) 2117 1.1 mrg { 2118 1.1 mrg return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, 2119 1.1 mrg (__v32hi) __B, 2120 1.1 mrg (__v32hi) 2121 1.3 mrg _mm512_setzero_si512 (), 2122 1.1 mrg (__mmask32) -1); 2123 1.1 mrg } 2124 1.1 mrg 2125 1.1 mrg extern __inline __m512i 2126 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2127 1.1 mrg _mm512_maskz_max_epu16 (__mmask32 __M, __m512i __A, __m512i __B) 2128 1.1 mrg { 2129 1.1 mrg return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, 2130 1.1 mrg (__v32hi) __B, 2131 1.1 mrg (__v32hi) 2132 1.3 mrg _mm512_setzero_si512 (), 2133 1.1 mrg (__mmask32) __M); 2134 1.1 mrg } 2135 1.1 mrg 2136 1.1 mrg extern __inline __m512i 2137 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2138 1.1 mrg _mm512_mask_max_epu16 (__m512i __W, __mmask32 __M, __m512i __A, 2139 1.1 mrg __m512i __B) 2140 1.1 mrg { 2141 1.1 mrg return (__m512i) __builtin_ia32_pmaxuw512_mask ((__v32hi) __A, 2142 1.1 mrg (__v32hi) __B, 2143 1.1 mrg (__v32hi) __W, 2144 1.1 mrg (__mmask32) __M); 2145 1.1 mrg } 2146 1.1 mrg 2147 1.1 mrg extern __inline __m512i 2148 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2149 1.1 mrg _mm512_sra_epi16 (__m512i __A, __m128i __B) 2150 1.1 mrg { 2151 1.1 mrg return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, 2152 1.1 mrg (__v8hi) __B, 2153 1.1 mrg (__v32hi) 2154 1.3 mrg _mm512_setzero_si512 (), 2155 1.1 mrg (__mmask32) -1); 2156 1.1 mrg } 2157 1.1 mrg 2158 1.1 mrg extern __inline __m512i 2159 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2160 1.1 mrg _mm512_mask_sra_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2161 1.1 mrg __m128i __B) 2162 1.1 mrg { 2163 1.1 mrg return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, 2164 1.1 mrg (__v8hi) __B, 2165 1.1 mrg (__v32hi) __W, 2166 1.1 mrg (__mmask32) __U); 2167 1.1 mrg } 2168 1.1 mrg 2169 1.1 mrg extern __inline __m512i 2170 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2171 1.1 mrg _mm512_maskz_sra_epi16 (__mmask32 __U, __m512i __A, __m128i __B) 2172 1.1 mrg { 2173 1.1 mrg return (__m512i) __builtin_ia32_psraw512_mask ((__v32hi) __A, 2174 1.1 mrg (__v8hi) __B, 2175 1.1 mrg (__v32hi) 2176 1.3 mrg _mm512_setzero_si512 (), 2177 1.1 mrg (__mmask32) __U); 2178 1.1 mrg } 2179 1.1 mrg 2180 1.1 mrg extern __inline __m512i 2181 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2182 1.1 mrg _mm512_srav_epi16 (__m512i __A, __m512i __B) 2183 1.1 mrg { 2184 1.1 mrg return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, 2185 1.1 mrg (__v32hi) __B, 2186 1.1 mrg (__v32hi) 2187 1.3 mrg _mm512_setzero_si512 (), 2188 1.1 mrg (__mmask32) -1); 2189 1.1 mrg } 2190 1.1 mrg 2191 1.1 mrg extern __inline __m512i 2192 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2193 1.1 mrg _mm512_mask_srav_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2194 1.1 mrg __m512i __B) 2195 1.1 mrg { 2196 1.1 mrg return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, 2197 1.1 mrg (__v32hi) __B, 2198 1.1 mrg (__v32hi) __W, 2199 1.1 mrg (__mmask32) __U); 2200 1.1 mrg } 2201 1.1 mrg 2202 1.1 mrg extern __inline __m512i 2203 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2204 1.1 mrg _mm512_maskz_srav_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 2205 1.1 mrg { 2206 1.1 mrg return (__m512i) __builtin_ia32_psrav32hi_mask ((__v32hi) __A, 2207 1.1 mrg (__v32hi) __B, 2208 1.1 mrg (__v32hi) 2209 1.3 mrg _mm512_setzero_si512 (), 2210 1.1 mrg (__mmask32) __U); 2211 1.1 mrg } 2212 1.1 mrg 2213 1.1 mrg extern __inline __m512i 2214 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2215 1.1 mrg _mm512_srlv_epi16 (__m512i __A, __m512i __B) 2216 1.1 mrg { 2217 1.1 mrg return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, 2218 1.1 mrg (__v32hi) __B, 2219 1.1 mrg (__v32hi) 2220 1.3 mrg _mm512_setzero_si512 (), 2221 1.1 mrg (__mmask32) -1); 2222 1.1 mrg } 2223 1.1 mrg 2224 1.1 mrg extern __inline __m512i 2225 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2226 1.1 mrg _mm512_mask_srlv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2227 1.1 mrg __m512i __B) 2228 1.1 mrg { 2229 1.1 mrg return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, 2230 1.1 mrg (__v32hi) __B, 2231 1.1 mrg (__v32hi) __W, 2232 1.1 mrg (__mmask32) __U); 2233 1.1 mrg } 2234 1.1 mrg 2235 1.1 mrg extern __inline __m512i 2236 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2237 1.1 mrg _mm512_maskz_srlv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 2238 1.1 mrg { 2239 1.1 mrg return (__m512i) __builtin_ia32_psrlv32hi_mask ((__v32hi) __A, 2240 1.1 mrg (__v32hi) __B, 2241 1.1 mrg (__v32hi) 2242 1.3 mrg _mm512_setzero_si512 (), 2243 1.1 mrg (__mmask32) __U); 2244 1.1 mrg } 2245 1.1 mrg 2246 1.1 mrg extern __inline __m512i 2247 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2248 1.1 mrg _mm512_sllv_epi16 (__m512i __A, __m512i __B) 2249 1.1 mrg { 2250 1.1 mrg return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, 2251 1.1 mrg (__v32hi) __B, 2252 1.1 mrg (__v32hi) 2253 1.3 mrg _mm512_setzero_si512 (), 2254 1.1 mrg (__mmask32) -1); 2255 1.1 mrg } 2256 1.1 mrg 2257 1.1 mrg extern __inline __m512i 2258 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2259 1.1 mrg _mm512_mask_sllv_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2260 1.1 mrg __m512i __B) 2261 1.1 mrg { 2262 1.1 mrg return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, 2263 1.1 mrg (__v32hi) __B, 2264 1.1 mrg (__v32hi) __W, 2265 1.1 mrg (__mmask32) __U); 2266 1.1 mrg } 2267 1.1 mrg 2268 1.1 mrg extern __inline __m512i 2269 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2270 1.1 mrg _mm512_maskz_sllv_epi16 (__mmask32 __U, __m512i __A, __m512i __B) 2271 1.1 mrg { 2272 1.1 mrg return (__m512i) __builtin_ia32_psllv32hi_mask ((__v32hi) __A, 2273 1.1 mrg (__v32hi) __B, 2274 1.1 mrg (__v32hi) 2275 1.3 mrg _mm512_setzero_si512 (), 2276 1.1 mrg (__mmask32) __U); 2277 1.1 mrg } 2278 1.1 mrg 2279 1.1 mrg extern __inline __m512i 2280 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2281 1.1 mrg _mm512_mask_packs_epi16 (__m512i __W, __mmask64 __M, __m512i __A, 2282 1.1 mrg __m512i __B) 2283 1.1 mrg { 2284 1.1 mrg return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, 2285 1.1 mrg (__v32hi) __B, 2286 1.1 mrg (__v64qi) __W, 2287 1.1 mrg (__mmask64) __M); 2288 1.1 mrg } 2289 1.1 mrg 2290 1.1 mrg extern __inline __m512i 2291 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2292 1.1 mrg _mm512_maskz_packs_epi16 (__mmask64 __M, __m512i __A, __m512i __B) 2293 1.1 mrg { 2294 1.1 mrg return (__m512i) __builtin_ia32_packsswb512_mask ((__v32hi) __A, 2295 1.1 mrg (__v32hi) __B, 2296 1.1 mrg (__v64qi) 2297 1.3 mrg _mm512_setzero_si512 (), 2298 1.1 mrg __M); 2299 1.1 mrg } 2300 1.1 mrg 2301 1.1 mrg extern __inline __m512i 2302 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2303 1.1 mrg _mm512_packus_epi16 (__m512i __A, __m512i __B) 2304 1.1 mrg { 2305 1.1 mrg return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, 2306 1.1 mrg (__v32hi) __B, 2307 1.1 mrg (__v64qi) 2308 1.3 mrg _mm512_setzero_si512 (), 2309 1.1 mrg (__mmask64) -1); 2310 1.1 mrg } 2311 1.1 mrg 2312 1.1 mrg extern __inline __m512i 2313 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2314 1.1 mrg _mm512_mask_packus_epi16 (__m512i __W, __mmask64 __M, __m512i __A, 2315 1.1 mrg __m512i __B) 2316 1.1 mrg { 2317 1.1 mrg return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, 2318 1.1 mrg (__v32hi) __B, 2319 1.1 mrg (__v64qi) __W, 2320 1.1 mrg (__mmask64) __M); 2321 1.1 mrg } 2322 1.1 mrg 2323 1.1 mrg extern __inline __m512i 2324 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2325 1.1 mrg _mm512_maskz_packus_epi16 (__mmask64 __M, __m512i __A, __m512i __B) 2326 1.1 mrg { 2327 1.1 mrg return (__m512i) __builtin_ia32_packuswb512_mask ((__v32hi) __A, 2328 1.1 mrg (__v32hi) __B, 2329 1.1 mrg (__v64qi) 2330 1.3 mrg _mm512_setzero_si512 (), 2331 1.1 mrg (__mmask64) __M); 2332 1.1 mrg } 2333 1.1 mrg 2334 1.1 mrg extern __inline __m512i 2335 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2336 1.1 mrg _mm512_abs_epi8 (__m512i __A) 2337 1.1 mrg { 2338 1.1 mrg return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, 2339 1.1 mrg (__v64qi) 2340 1.3 mrg _mm512_setzero_si512 (), 2341 1.1 mrg (__mmask64) -1); 2342 1.1 mrg } 2343 1.1 mrg 2344 1.1 mrg extern __inline __m512i 2345 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2346 1.1 mrg _mm512_mask_abs_epi8 (__m512i __W, __mmask64 __U, __m512i __A) 2347 1.1 mrg { 2348 1.1 mrg return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, 2349 1.1 mrg (__v64qi) __W, 2350 1.1 mrg (__mmask64) __U); 2351 1.1 mrg } 2352 1.1 mrg 2353 1.1 mrg extern __inline __m512i 2354 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2355 1.1 mrg _mm512_maskz_abs_epi8 (__mmask64 __U, __m512i __A) 2356 1.1 mrg { 2357 1.1 mrg return (__m512i) __builtin_ia32_pabsb512_mask ((__v64qi) __A, 2358 1.1 mrg (__v64qi) 2359 1.3 mrg _mm512_setzero_si512 (), 2360 1.1 mrg (__mmask64) __U); 2361 1.1 mrg } 2362 1.1 mrg 2363 1.1 mrg extern __inline __m512i 2364 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2365 1.1 mrg _mm512_abs_epi16 (__m512i __A) 2366 1.1 mrg { 2367 1.1 mrg return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, 2368 1.1 mrg (__v32hi) 2369 1.3 mrg _mm512_setzero_si512 (), 2370 1.1 mrg (__mmask32) -1); 2371 1.1 mrg } 2372 1.1 mrg 2373 1.1 mrg extern __inline __m512i 2374 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2375 1.1 mrg _mm512_mask_abs_epi16 (__m512i __W, __mmask32 __U, __m512i __A) 2376 1.1 mrg { 2377 1.1 mrg return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, 2378 1.1 mrg (__v32hi) __W, 2379 1.1 mrg (__mmask32) __U); 2380 1.1 mrg } 2381 1.1 mrg 2382 1.1 mrg extern __inline __m512i 2383 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2384 1.1 mrg _mm512_maskz_abs_epi16 (__mmask32 __U, __m512i __A) 2385 1.1 mrg { 2386 1.1 mrg return (__m512i) __builtin_ia32_pabsw512_mask ((__v32hi) __A, 2387 1.1 mrg (__v32hi) 2388 1.3 mrg _mm512_setzero_si512 (), 2389 1.1 mrg (__mmask32) __U); 2390 1.1 mrg } 2391 1.1 mrg 2392 1.1 mrg extern __inline __mmask64 2393 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2394 1.1 mrg _mm512_mask_cmpneq_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2395 1.1 mrg { 2396 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2397 1.1 mrg (__v64qi) __Y, 4, 2398 1.1 mrg (__mmask64) __M); 2399 1.1 mrg } 2400 1.1 mrg 2401 1.1 mrg extern __inline __mmask64 2402 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2403 1.1 mrg _mm512_mask_cmplt_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2404 1.1 mrg { 2405 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2406 1.1 mrg (__v64qi) __Y, 1, 2407 1.1 mrg (__mmask64) __M); 2408 1.1 mrg } 2409 1.1 mrg 2410 1.1 mrg extern __inline __mmask64 2411 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2412 1.1 mrg _mm512_mask_cmpge_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2413 1.1 mrg { 2414 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2415 1.1 mrg (__v64qi) __Y, 5, 2416 1.1 mrg (__mmask64) __M); 2417 1.1 mrg } 2418 1.1 mrg 2419 1.1 mrg extern __inline __mmask64 2420 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2421 1.1 mrg _mm512_mask_cmple_epu8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2422 1.1 mrg { 2423 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2424 1.1 mrg (__v64qi) __Y, 2, 2425 1.1 mrg (__mmask64) __M); 2426 1.1 mrg } 2427 1.1 mrg 2428 1.1 mrg extern __inline __mmask32 2429 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2430 1.1 mrg _mm512_mask_cmpneq_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2431 1.1 mrg { 2432 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2433 1.1 mrg (__v32hi) __Y, 4, 2434 1.1 mrg (__mmask32) __M); 2435 1.1 mrg } 2436 1.1 mrg 2437 1.1 mrg extern __inline __mmask32 2438 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2439 1.1 mrg _mm512_mask_cmplt_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2440 1.1 mrg { 2441 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2442 1.1 mrg (__v32hi) __Y, 1, 2443 1.1 mrg (__mmask32) __M); 2444 1.1 mrg } 2445 1.1 mrg 2446 1.1 mrg extern __inline __mmask32 2447 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2448 1.1 mrg _mm512_mask_cmpge_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2449 1.1 mrg { 2450 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2451 1.1 mrg (__v32hi) __Y, 5, 2452 1.1 mrg (__mmask32) __M); 2453 1.1 mrg } 2454 1.1 mrg 2455 1.1 mrg extern __inline __mmask32 2456 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2457 1.1 mrg _mm512_mask_cmple_epu16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2458 1.1 mrg { 2459 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2460 1.1 mrg (__v32hi) __Y, 2, 2461 1.1 mrg (__mmask32) __M); 2462 1.1 mrg } 2463 1.1 mrg 2464 1.1 mrg extern __inline __mmask64 2465 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2466 1.1 mrg _mm512_mask_cmpneq_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2467 1.1 mrg { 2468 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2469 1.1 mrg (__v64qi) __Y, 4, 2470 1.1 mrg (__mmask64) __M); 2471 1.1 mrg } 2472 1.1 mrg 2473 1.1 mrg extern __inline __mmask64 2474 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2475 1.1 mrg _mm512_mask_cmplt_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2476 1.1 mrg { 2477 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2478 1.1 mrg (__v64qi) __Y, 1, 2479 1.1 mrg (__mmask64) __M); 2480 1.1 mrg } 2481 1.1 mrg 2482 1.1 mrg extern __inline __mmask64 2483 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2484 1.1 mrg _mm512_mask_cmpge_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2485 1.1 mrg { 2486 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2487 1.1 mrg (__v64qi) __Y, 5, 2488 1.1 mrg (__mmask64) __M); 2489 1.1 mrg } 2490 1.1 mrg 2491 1.1 mrg extern __inline __mmask64 2492 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2493 1.1 mrg _mm512_mask_cmple_epi8_mask (__mmask64 __M, __m512i __X, __m512i __Y) 2494 1.1 mrg { 2495 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2496 1.1 mrg (__v64qi) __Y, 2, 2497 1.1 mrg (__mmask64) __M); 2498 1.1 mrg } 2499 1.1 mrg 2500 1.1 mrg extern __inline __mmask32 2501 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2502 1.1 mrg _mm512_mask_cmpneq_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2503 1.1 mrg { 2504 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2505 1.1 mrg (__v32hi) __Y, 4, 2506 1.1 mrg (__mmask32) __M); 2507 1.1 mrg } 2508 1.1 mrg 2509 1.1 mrg extern __inline __mmask32 2510 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2511 1.1 mrg _mm512_mask_cmplt_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2512 1.1 mrg { 2513 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2514 1.1 mrg (__v32hi) __Y, 1, 2515 1.1 mrg (__mmask32) __M); 2516 1.1 mrg } 2517 1.1 mrg 2518 1.1 mrg extern __inline __mmask32 2519 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2520 1.1 mrg _mm512_mask_cmpge_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2521 1.1 mrg { 2522 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2523 1.1 mrg (__v32hi) __Y, 5, 2524 1.1 mrg (__mmask32) __M); 2525 1.1 mrg } 2526 1.1 mrg 2527 1.1 mrg extern __inline __mmask32 2528 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2529 1.1 mrg _mm512_mask_cmple_epi16_mask (__mmask32 __M, __m512i __X, __m512i __Y) 2530 1.1 mrg { 2531 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2532 1.1 mrg (__v32hi) __Y, 2, 2533 1.1 mrg (__mmask32) __M); 2534 1.1 mrg } 2535 1.1 mrg 2536 1.1 mrg extern __inline __mmask64 2537 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2538 1.1 mrg _mm512_cmpneq_epu8_mask (__m512i __X, __m512i __Y) 2539 1.1 mrg { 2540 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2541 1.1 mrg (__v64qi) __Y, 4, 2542 1.3 mrg (__mmask64) -1); 2543 1.1 mrg } 2544 1.1 mrg 2545 1.1 mrg extern __inline __mmask64 2546 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2547 1.1 mrg _mm512_cmplt_epu8_mask (__m512i __X, __m512i __Y) 2548 1.1 mrg { 2549 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2550 1.1 mrg (__v64qi) __Y, 1, 2551 1.3 mrg (__mmask64) -1); 2552 1.1 mrg } 2553 1.1 mrg 2554 1.1 mrg extern __inline __mmask64 2555 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2556 1.1 mrg _mm512_cmpge_epu8_mask (__m512i __X, __m512i __Y) 2557 1.1 mrg { 2558 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2559 1.1 mrg (__v64qi) __Y, 5, 2560 1.3 mrg (__mmask64) -1); 2561 1.1 mrg } 2562 1.1 mrg 2563 1.1 mrg extern __inline __mmask64 2564 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2565 1.1 mrg _mm512_cmple_epu8_mask (__m512i __X, __m512i __Y) 2566 1.1 mrg { 2567 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 2568 1.1 mrg (__v64qi) __Y, 2, 2569 1.3 mrg (__mmask64) -1); 2570 1.1 mrg } 2571 1.1 mrg 2572 1.1 mrg extern __inline __mmask32 2573 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2574 1.1 mrg _mm512_cmpneq_epu16_mask (__m512i __X, __m512i __Y) 2575 1.1 mrg { 2576 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2577 1.1 mrg (__v32hi) __Y, 4, 2578 1.3 mrg (__mmask32) -1); 2579 1.1 mrg } 2580 1.1 mrg 2581 1.1 mrg extern __inline __mmask32 2582 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2583 1.1 mrg _mm512_cmplt_epu16_mask (__m512i __X, __m512i __Y) 2584 1.1 mrg { 2585 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2586 1.1 mrg (__v32hi) __Y, 1, 2587 1.3 mrg (__mmask32) -1); 2588 1.1 mrg } 2589 1.1 mrg 2590 1.1 mrg extern __inline __mmask32 2591 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2592 1.1 mrg _mm512_cmpge_epu16_mask (__m512i __X, __m512i __Y) 2593 1.1 mrg { 2594 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2595 1.1 mrg (__v32hi) __Y, 5, 2596 1.3 mrg (__mmask32) -1); 2597 1.1 mrg } 2598 1.1 mrg 2599 1.1 mrg extern __inline __mmask32 2600 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2601 1.1 mrg _mm512_cmple_epu16_mask (__m512i __X, __m512i __Y) 2602 1.1 mrg { 2603 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 2604 1.1 mrg (__v32hi) __Y, 2, 2605 1.3 mrg (__mmask32) -1); 2606 1.1 mrg } 2607 1.1 mrg 2608 1.1 mrg extern __inline __mmask64 2609 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2610 1.1 mrg _mm512_cmpneq_epi8_mask (__m512i __X, __m512i __Y) 2611 1.1 mrg { 2612 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2613 1.1 mrg (__v64qi) __Y, 4, 2614 1.3 mrg (__mmask64) -1); 2615 1.1 mrg } 2616 1.1 mrg 2617 1.1 mrg extern __inline __mmask64 2618 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2619 1.1 mrg _mm512_cmplt_epi8_mask (__m512i __X, __m512i __Y) 2620 1.1 mrg { 2621 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2622 1.1 mrg (__v64qi) __Y, 1, 2623 1.3 mrg (__mmask64) -1); 2624 1.1 mrg } 2625 1.1 mrg 2626 1.1 mrg extern __inline __mmask64 2627 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2628 1.1 mrg _mm512_cmpge_epi8_mask (__m512i __X, __m512i __Y) 2629 1.1 mrg { 2630 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2631 1.1 mrg (__v64qi) __Y, 5, 2632 1.3 mrg (__mmask64) -1); 2633 1.1 mrg } 2634 1.1 mrg 2635 1.1 mrg extern __inline __mmask64 2636 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2637 1.1 mrg _mm512_cmple_epi8_mask (__m512i __X, __m512i __Y) 2638 1.1 mrg { 2639 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 2640 1.1 mrg (__v64qi) __Y, 2, 2641 1.3 mrg (__mmask64) -1); 2642 1.1 mrg } 2643 1.1 mrg 2644 1.1 mrg extern __inline __mmask32 2645 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2646 1.1 mrg _mm512_cmpneq_epi16_mask (__m512i __X, __m512i __Y) 2647 1.1 mrg { 2648 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2649 1.1 mrg (__v32hi) __Y, 4, 2650 1.3 mrg (__mmask32) -1); 2651 1.1 mrg } 2652 1.1 mrg 2653 1.1 mrg extern __inline __mmask32 2654 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2655 1.1 mrg _mm512_cmplt_epi16_mask (__m512i __X, __m512i __Y) 2656 1.1 mrg { 2657 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2658 1.1 mrg (__v32hi) __Y, 1, 2659 1.3 mrg (__mmask32) -1); 2660 1.1 mrg } 2661 1.1 mrg 2662 1.1 mrg extern __inline __mmask32 2663 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2664 1.1 mrg _mm512_cmpge_epi16_mask (__m512i __X, __m512i __Y) 2665 1.1 mrg { 2666 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2667 1.1 mrg (__v32hi) __Y, 5, 2668 1.3 mrg (__mmask32) -1); 2669 1.1 mrg } 2670 1.1 mrg 2671 1.1 mrg extern __inline __mmask32 2672 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2673 1.1 mrg _mm512_cmple_epi16_mask (__m512i __X, __m512i __Y) 2674 1.1 mrg { 2675 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 2676 1.1 mrg (__v32hi) __Y, 2, 2677 1.3 mrg (__mmask32) -1); 2678 1.3 mrg } 2679 1.3 mrg 2680 1.3 mrg extern __inline __m512i 2681 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2682 1.3 mrg _mm512_packs_epi32 (__m512i __A, __m512i __B) 2683 1.3 mrg { 2684 1.3 mrg return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, 2685 1.3 mrg (__v16si) __B, 2686 1.3 mrg (__v32hi) 2687 1.3 mrg _mm512_setzero_si512 (), 2688 1.3 mrg (__mmask32) -1); 2689 1.3 mrg } 2690 1.3 mrg 2691 1.3 mrg extern __inline __m512i 2692 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2693 1.3 mrg _mm512_maskz_packs_epi32 (__mmask32 __M, __m512i __A, __m512i __B) 2694 1.3 mrg { 2695 1.3 mrg return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, 2696 1.3 mrg (__v16si) __B, 2697 1.3 mrg (__v32hi) 2698 1.3 mrg _mm512_setzero_si512 (), 2699 1.3 mrg __M); 2700 1.3 mrg } 2701 1.3 mrg 2702 1.3 mrg extern __inline __m512i 2703 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2704 1.3 mrg _mm512_mask_packs_epi32 (__m512i __W, __mmask32 __M, __m512i __A, 2705 1.3 mrg __m512i __B) 2706 1.3 mrg { 2707 1.3 mrg return (__m512i) __builtin_ia32_packssdw512_mask ((__v16si) __A, 2708 1.3 mrg (__v16si) __B, 2709 1.3 mrg (__v32hi) __W, 2710 1.3 mrg __M); 2711 1.3 mrg } 2712 1.3 mrg 2713 1.3 mrg extern __inline __m512i 2714 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2715 1.3 mrg _mm512_packus_epi32 (__m512i __A, __m512i __B) 2716 1.3 mrg { 2717 1.3 mrg return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, 2718 1.3 mrg (__v16si) __B, 2719 1.3 mrg (__v32hi) 2720 1.3 mrg _mm512_setzero_si512 (), 2721 1.3 mrg (__mmask32) -1); 2722 1.3 mrg } 2723 1.3 mrg 2724 1.3 mrg extern __inline __m512i 2725 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2726 1.3 mrg _mm512_maskz_packus_epi32 (__mmask32 __M, __m512i __A, __m512i __B) 2727 1.3 mrg { 2728 1.3 mrg return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, 2729 1.3 mrg (__v16si) __B, 2730 1.3 mrg (__v32hi) 2731 1.3 mrg _mm512_setzero_si512 (), 2732 1.3 mrg __M); 2733 1.3 mrg } 2734 1.3 mrg 2735 1.3 mrg extern __inline __m512i 2736 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2737 1.3 mrg _mm512_mask_packus_epi32 (__m512i __W, __mmask32 __M, __m512i __A, 2738 1.3 mrg __m512i __B) 2739 1.3 mrg { 2740 1.3 mrg return (__m512i) __builtin_ia32_packusdw512_mask ((__v16si) __A, 2741 1.3 mrg (__v16si) __B, 2742 1.3 mrg (__v32hi) __W, 2743 1.3 mrg __M); 2744 1.1 mrg } 2745 1.1 mrg 2746 1.1 mrg #ifdef __OPTIMIZE__ 2747 1.3 mrg extern __inline __mmask32 2748 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2749 1.3 mrg _kshiftli_mask32 (__mmask32 __A, unsigned int __B) 2750 1.3 mrg { 2751 1.3 mrg return (__mmask32) __builtin_ia32_kshiftlisi ((__mmask32) __A, 2752 1.3 mrg (__mmask8) __B); 2753 1.3 mrg } 2754 1.3 mrg 2755 1.3 mrg extern __inline __mmask64 2756 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2757 1.3 mrg _kshiftli_mask64 (__mmask64 __A, unsigned int __B) 2758 1.3 mrg { 2759 1.3 mrg return (__mmask64) __builtin_ia32_kshiftlidi ((__mmask64) __A, 2760 1.3 mrg (__mmask8) __B); 2761 1.3 mrg } 2762 1.3 mrg 2763 1.3 mrg extern __inline __mmask32 2764 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2765 1.3 mrg _kshiftri_mask32 (__mmask32 __A, unsigned int __B) 2766 1.3 mrg { 2767 1.3 mrg return (__mmask32) __builtin_ia32_kshiftrisi ((__mmask32) __A, 2768 1.3 mrg (__mmask8) __B); 2769 1.3 mrg } 2770 1.3 mrg 2771 1.3 mrg extern __inline __mmask64 2772 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2773 1.3 mrg _kshiftri_mask64 (__mmask64 __A, unsigned int __B) 2774 1.3 mrg { 2775 1.3 mrg return (__mmask64) __builtin_ia32_kshiftridi ((__mmask64) __A, 2776 1.3 mrg (__mmask8) __B); 2777 1.3 mrg } 2778 1.3 mrg 2779 1.1 mrg extern __inline __m512i 2780 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2781 1.1 mrg _mm512_alignr_epi8 (__m512i __A, __m512i __B, const int __N) 2782 1.1 mrg { 2783 1.1 mrg return (__m512i) __builtin_ia32_palignr512 ((__v8di) __A, 2784 1.1 mrg (__v8di) __B, __N * 8); 2785 1.1 mrg } 2786 1.1 mrg 2787 1.1 mrg extern __inline __m512i 2788 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2789 1.1 mrg _mm512_mask_alignr_epi8 (__m512i __W, __mmask64 __U, __m512i __A, 2790 1.1 mrg __m512i __B, const int __N) 2791 1.1 mrg { 2792 1.1 mrg return (__m512i) __builtin_ia32_palignr512_mask ((__v8di) __A, 2793 1.1 mrg (__v8di) __B, 2794 1.1 mrg __N * 8, 2795 1.1 mrg (__v8di) __W, 2796 1.1 mrg (__mmask64) __U); 2797 1.1 mrg } 2798 1.1 mrg 2799 1.1 mrg extern __inline __m512i 2800 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2801 1.1 mrg _mm512_maskz_alignr_epi8 (__mmask64 __U, __m512i __A, __m512i __B, 2802 1.1 mrg const int __N) 2803 1.1 mrg { 2804 1.1 mrg return (__m512i) __builtin_ia32_palignr512_mask ((__v8di) __A, 2805 1.1 mrg (__v8di) __B, 2806 1.1 mrg __N * 8, 2807 1.1 mrg (__v8di) 2808 1.1 mrg _mm512_setzero_si512 (), 2809 1.1 mrg (__mmask64) __U); 2810 1.1 mrg } 2811 1.1 mrg 2812 1.1 mrg extern __inline __m512i 2813 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2814 1.1 mrg _mm512_dbsad_epu8 (__m512i __A, __m512i __B, const int __imm) 2815 1.1 mrg { 2816 1.1 mrg return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, 2817 1.1 mrg (__v64qi) __B, 2818 1.1 mrg __imm, 2819 1.1 mrg (__v32hi) 2820 1.3 mrg _mm512_setzero_si512 (), 2821 1.1 mrg (__mmask32) -1); 2822 1.1 mrg } 2823 1.1 mrg 2824 1.1 mrg extern __inline __m512i 2825 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2826 1.1 mrg _mm512_mask_dbsad_epu8 (__m512i __W, __mmask32 __U, __m512i __A, 2827 1.1 mrg __m512i __B, const int __imm) 2828 1.1 mrg { 2829 1.1 mrg return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, 2830 1.1 mrg (__v64qi) __B, 2831 1.1 mrg __imm, 2832 1.1 mrg (__v32hi) __W, 2833 1.1 mrg (__mmask32) __U); 2834 1.1 mrg } 2835 1.1 mrg 2836 1.1 mrg extern __inline __m512i 2837 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2838 1.1 mrg _mm512_maskz_dbsad_epu8 (__mmask32 __U, __m512i __A, __m512i __B, 2839 1.1 mrg const int __imm) 2840 1.1 mrg { 2841 1.1 mrg return (__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi) __A, 2842 1.1 mrg (__v64qi) __B, 2843 1.1 mrg __imm, 2844 1.1 mrg (__v32hi) 2845 1.3 mrg _mm512_setzero_si512 (), 2846 1.1 mrg (__mmask32) __U); 2847 1.1 mrg } 2848 1.1 mrg 2849 1.1 mrg extern __inline __m512i 2850 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2851 1.1 mrg _mm512_srli_epi16 (__m512i __A, const int __imm) 2852 1.1 mrg { 2853 1.1 mrg return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, 2854 1.1 mrg (__v32hi) 2855 1.3 mrg _mm512_setzero_si512 (), 2856 1.1 mrg (__mmask32) -1); 2857 1.1 mrg } 2858 1.1 mrg 2859 1.1 mrg extern __inline __m512i 2860 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2861 1.1 mrg _mm512_mask_srli_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2862 1.1 mrg const int __imm) 2863 1.1 mrg { 2864 1.1 mrg return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, 2865 1.1 mrg (__v32hi) __W, 2866 1.1 mrg (__mmask32) __U); 2867 1.1 mrg } 2868 1.1 mrg 2869 1.1 mrg extern __inline __m512i 2870 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2871 1.1 mrg _mm512_maskz_srli_epi16 (__mmask32 __U, __m512i __A, const int __imm) 2872 1.1 mrg { 2873 1.1 mrg return (__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi) __A, __imm, 2874 1.1 mrg (__v32hi) 2875 1.3 mrg _mm512_setzero_si512 (), 2876 1.1 mrg (__mmask32) __U); 2877 1.1 mrg } 2878 1.1 mrg 2879 1.1 mrg extern __inline __m512i 2880 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2881 1.1 mrg _mm512_slli_epi16 (__m512i __A, const int __B) 2882 1.1 mrg { 2883 1.1 mrg return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, 2884 1.1 mrg (__v32hi) 2885 1.3 mrg _mm512_setzero_si512 (), 2886 1.1 mrg (__mmask32) -1); 2887 1.1 mrg } 2888 1.1 mrg 2889 1.1 mrg extern __inline __m512i 2890 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2891 1.1 mrg _mm512_mask_slli_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2892 1.1 mrg const int __B) 2893 1.1 mrg { 2894 1.1 mrg return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, 2895 1.1 mrg (__v32hi) __W, 2896 1.1 mrg (__mmask32) __U); 2897 1.1 mrg } 2898 1.1 mrg 2899 1.1 mrg extern __inline __m512i 2900 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2901 1.1 mrg _mm512_maskz_slli_epi16 (__mmask32 __U, __m512i __A, const int __B) 2902 1.1 mrg { 2903 1.1 mrg return (__m512i) __builtin_ia32_psllwi512_mask ((__v32hi) __A, __B, 2904 1.1 mrg (__v32hi) 2905 1.3 mrg _mm512_setzero_si512 (), 2906 1.1 mrg (__mmask32) __U); 2907 1.1 mrg } 2908 1.1 mrg 2909 1.1 mrg extern __inline __m512i 2910 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2911 1.1 mrg _mm512_shufflehi_epi16 (__m512i __A, const int __imm) 2912 1.1 mrg { 2913 1.1 mrg return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, 2914 1.1 mrg __imm, 2915 1.1 mrg (__v32hi) 2916 1.3 mrg _mm512_setzero_si512 (), 2917 1.1 mrg (__mmask32) -1); 2918 1.1 mrg } 2919 1.1 mrg 2920 1.1 mrg extern __inline __m512i 2921 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2922 1.1 mrg _mm512_mask_shufflehi_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2923 1.1 mrg const int __imm) 2924 1.1 mrg { 2925 1.1 mrg return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, 2926 1.1 mrg __imm, 2927 1.1 mrg (__v32hi) __W, 2928 1.1 mrg (__mmask32) __U); 2929 1.1 mrg } 2930 1.1 mrg 2931 1.1 mrg extern __inline __m512i 2932 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2933 1.1 mrg _mm512_maskz_shufflehi_epi16 (__mmask32 __U, __m512i __A, 2934 1.1 mrg const int __imm) 2935 1.1 mrg { 2936 1.1 mrg return (__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi) __A, 2937 1.1 mrg __imm, 2938 1.1 mrg (__v32hi) 2939 1.3 mrg _mm512_setzero_si512 (), 2940 1.1 mrg (__mmask32) __U); 2941 1.1 mrg } 2942 1.1 mrg 2943 1.1 mrg extern __inline __m512i 2944 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2945 1.1 mrg _mm512_shufflelo_epi16 (__m512i __A, const int __imm) 2946 1.1 mrg { 2947 1.1 mrg return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, 2948 1.1 mrg __imm, 2949 1.1 mrg (__v32hi) 2950 1.3 mrg _mm512_setzero_si512 (), 2951 1.1 mrg (__mmask32) -1); 2952 1.1 mrg } 2953 1.1 mrg 2954 1.1 mrg extern __inline __m512i 2955 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2956 1.1 mrg _mm512_mask_shufflelo_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2957 1.1 mrg const int __imm) 2958 1.1 mrg { 2959 1.1 mrg return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, 2960 1.1 mrg __imm, 2961 1.1 mrg (__v32hi) __W, 2962 1.1 mrg (__mmask32) __U); 2963 1.1 mrg } 2964 1.1 mrg 2965 1.1 mrg extern __inline __m512i 2966 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2967 1.1 mrg _mm512_maskz_shufflelo_epi16 (__mmask32 __U, __m512i __A, 2968 1.1 mrg const int __imm) 2969 1.1 mrg { 2970 1.1 mrg return (__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi) __A, 2971 1.1 mrg __imm, 2972 1.1 mrg (__v32hi) 2973 1.3 mrg _mm512_setzero_si512 (), 2974 1.1 mrg (__mmask32) __U); 2975 1.1 mrg } 2976 1.1 mrg 2977 1.1 mrg extern __inline __m512i 2978 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2979 1.1 mrg _mm512_srai_epi16 (__m512i __A, const int __imm) 2980 1.1 mrg { 2981 1.1 mrg return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, 2982 1.1 mrg (__v32hi) 2983 1.3 mrg _mm512_setzero_si512 (), 2984 1.1 mrg (__mmask32) -1); 2985 1.1 mrg } 2986 1.1 mrg 2987 1.1 mrg extern __inline __m512i 2988 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2989 1.1 mrg _mm512_mask_srai_epi16 (__m512i __W, __mmask32 __U, __m512i __A, 2990 1.1 mrg const int __imm) 2991 1.1 mrg { 2992 1.1 mrg return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, 2993 1.1 mrg (__v32hi) __W, 2994 1.1 mrg (__mmask32) __U); 2995 1.1 mrg } 2996 1.1 mrg 2997 1.1 mrg extern __inline __m512i 2998 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2999 1.1 mrg _mm512_maskz_srai_epi16 (__mmask32 __U, __m512i __A, const int __imm) 3000 1.1 mrg { 3001 1.1 mrg return (__m512i) __builtin_ia32_psrawi512_mask ((__v32hi) __A, __imm, 3002 1.1 mrg (__v32hi) 3003 1.3 mrg _mm512_setzero_si512 (), 3004 1.1 mrg (__mmask32) __U); 3005 1.1 mrg } 3006 1.1 mrg 3007 1.1 mrg extern __inline __m512i 3008 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3009 1.1 mrg _mm512_mask_blend_epi16 (__mmask32 __U, __m512i __A, __m512i __W) 3010 1.1 mrg { 3011 1.1 mrg return (__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) __A, 3012 1.1 mrg (__v32hi) __W, 3013 1.1 mrg (__mmask32) __U); 3014 1.1 mrg } 3015 1.1 mrg 3016 1.1 mrg extern __inline __m512i 3017 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3018 1.1 mrg _mm512_mask_blend_epi8 (__mmask64 __U, __m512i __A, __m512i __W) 3019 1.1 mrg { 3020 1.1 mrg return (__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) __A, 3021 1.1 mrg (__v64qi) __W, 3022 1.1 mrg (__mmask64) __U); 3023 1.1 mrg } 3024 1.1 mrg 3025 1.1 mrg extern __inline __mmask32 3026 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3027 1.1 mrg _mm512_mask_cmp_epi16_mask (__mmask32 __U, __m512i __X, __m512i __Y, 3028 1.1 mrg const int __P) 3029 1.1 mrg { 3030 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 3031 1.1 mrg (__v32hi) __Y, __P, 3032 1.1 mrg (__mmask32) __U); 3033 1.1 mrg } 3034 1.1 mrg 3035 1.1 mrg extern __inline __mmask32 3036 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3037 1.1 mrg _mm512_cmp_epi16_mask (__m512i __X, __m512i __Y, const int __P) 3038 1.1 mrg { 3039 1.1 mrg return (__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi) __X, 3040 1.1 mrg (__v32hi) __Y, __P, 3041 1.1 mrg (__mmask32) -1); 3042 1.1 mrg } 3043 1.1 mrg 3044 1.1 mrg extern __inline __mmask64 3045 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3046 1.4 mrg _mm512_mask_cmp_epi8_mask (__mmask64 __U, __m512i __X, __m512i __Y, 3047 1.1 mrg const int __P) 3048 1.1 mrg { 3049 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 3050 1.1 mrg (__v64qi) __Y, __P, 3051 1.1 mrg (__mmask64) __U); 3052 1.1 mrg } 3053 1.1 mrg 3054 1.1 mrg extern __inline __mmask64 3055 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3056 1.1 mrg _mm512_cmp_epi8_mask (__m512i __X, __m512i __Y, const int __P) 3057 1.1 mrg { 3058 1.1 mrg return (__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi) __X, 3059 1.1 mrg (__v64qi) __Y, __P, 3060 1.1 mrg (__mmask64) -1); 3061 1.1 mrg } 3062 1.1 mrg 3063 1.1 mrg extern __inline __mmask32 3064 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3065 1.1 mrg _mm512_mask_cmp_epu16_mask (__mmask32 __U, __m512i __X, __m512i __Y, 3066 1.1 mrg const int __P) 3067 1.1 mrg { 3068 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 3069 1.1 mrg (__v32hi) __Y, __P, 3070 1.1 mrg (__mmask32) __U); 3071 1.1 mrg } 3072 1.1 mrg 3073 1.1 mrg extern __inline __mmask32 3074 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3075 1.1 mrg _mm512_cmp_epu16_mask (__m512i __X, __m512i __Y, const int __P) 3076 1.1 mrg { 3077 1.1 mrg return (__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi) __X, 3078 1.1 mrg (__v32hi) __Y, __P, 3079 1.1 mrg (__mmask32) -1); 3080 1.1 mrg } 3081 1.1 mrg 3082 1.1 mrg extern __inline __mmask64 3083 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3084 1.4 mrg _mm512_mask_cmp_epu8_mask (__mmask64 __U, __m512i __X, __m512i __Y, 3085 1.1 mrg const int __P) 3086 1.1 mrg { 3087 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 3088 1.1 mrg (__v64qi) __Y, __P, 3089 1.1 mrg (__mmask64) __U); 3090 1.1 mrg } 3091 1.1 mrg 3092 1.1 mrg extern __inline __mmask64 3093 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3094 1.1 mrg _mm512_cmp_epu8_mask (__m512i __X, __m512i __Y, const int __P) 3095 1.1 mrg { 3096 1.1 mrg return (__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi) __X, 3097 1.1 mrg (__v64qi) __Y, __P, 3098 1.1 mrg (__mmask64) -1); 3099 1.1 mrg } 3100 1.1 mrg 3101 1.1 mrg extern __inline __m512i 3102 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3103 1.1 mrg _mm512_bslli_epi128 (__m512i __A, const int __N) 3104 1.1 mrg { 3105 1.1 mrg return (__m512i) __builtin_ia32_pslldq512 (__A, __N * 8); 3106 1.1 mrg } 3107 1.1 mrg 3108 1.1 mrg extern __inline __m512i 3109 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 3110 1.1 mrg _mm512_bsrli_epi128 (__m512i __A, const int __N) 3111 1.1 mrg { 3112 1.1 mrg return (__m512i) __builtin_ia32_psrldq512 (__A, __N * 8); 3113 1.1 mrg } 3114 1.1 mrg 3115 1.1 mrg #else 3116 1.3 mrg #define _kshiftli_mask32(X, Y) \ 3117 1.3 mrg ((__mmask32) __builtin_ia32_kshiftlisi ((__mmask32)(X), (__mmask8)(Y))) 3118 1.3 mrg 3119 1.3 mrg #define _kshiftli_mask64(X, Y) \ 3120 1.3 mrg ((__mmask64) __builtin_ia32_kshiftlidi ((__mmask64)(X), (__mmask8)(Y))) 3121 1.3 mrg 3122 1.3 mrg #define _kshiftri_mask32(X, Y) \ 3123 1.3 mrg ((__mmask32) __builtin_ia32_kshiftrisi ((__mmask32)(X), (__mmask8)(Y))) 3124 1.3 mrg 3125 1.3 mrg #define _kshiftri_mask64(X, Y) \ 3126 1.3 mrg ((__mmask64) __builtin_ia32_kshiftridi ((__mmask64)(X), (__mmask8)(Y))) 3127 1.3 mrg 3128 1.1 mrg #define _mm512_alignr_epi8(X, Y, N) \ 3129 1.1 mrg ((__m512i) __builtin_ia32_palignr512 ((__v8di)(__m512i)(X), \ 3130 1.1 mrg (__v8di)(__m512i)(Y), \ 3131 1.6 mrg (int)((N) * 8))) 3132 1.1 mrg 3133 1.1 mrg #define _mm512_mask_alignr_epi8(W, U, X, Y, N) \ 3134 1.1 mrg ((__m512i) __builtin_ia32_palignr512_mask ((__v8di)(__m512i)(X), \ 3135 1.6 mrg (__v8di)(__m512i)(Y), (int)((N) * 8), \ 3136 1.1 mrg (__v8di)(__m512i)(W), (__mmask64)(U))) 3137 1.1 mrg 3138 1.1 mrg #define _mm512_maskz_alignr_epi8(U, X, Y, N) \ 3139 1.1 mrg ((__m512i) __builtin_ia32_palignr512_mask ((__v8di)(__m512i)(X), \ 3140 1.6 mrg (__v8di)(__m512i)(Y), (int)((N) * 8), \ 3141 1.3 mrg (__v8di)(__m512i) \ 3142 1.3 mrg _mm512_setzero_si512 (), \ 3143 1.3 mrg (__mmask64)(U))) 3144 1.1 mrg 3145 1.1 mrg #define _mm512_dbsad_epu8(X, Y, C) \ 3146 1.1 mrg ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ 3147 1.1 mrg (__v64qi)(__m512i) (Y), (int) (C), \ 3148 1.3 mrg (__v32hi)(__m512i) \ 3149 1.3 mrg _mm512_setzero_si512 (), \ 3150 1.1 mrg (__mmask32)-1)) 3151 1.1 mrg 3152 1.1 mrg #define _mm512_mask_dbsad_epu8(W, U, X, Y, C) \ 3153 1.1 mrg ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ 3154 1.1 mrg (__v64qi)(__m512i) (Y), (int) (C), \ 3155 1.1 mrg (__v32hi)(__m512i)(W), \ 3156 1.1 mrg (__mmask32)(U))) 3157 1.1 mrg 3158 1.1 mrg #define _mm512_maskz_dbsad_epu8(U, X, Y, C) \ 3159 1.1 mrg ((__m512i) __builtin_ia32_dbpsadbw512_mask ((__v64qi)(__m512i) (X), \ 3160 1.1 mrg (__v64qi)(__m512i) (Y), (int) (C), \ 3161 1.3 mrg (__v32hi)(__m512i) \ 3162 1.3 mrg _mm512_setzero_si512 (), \ 3163 1.1 mrg (__mmask32)(U))) 3164 1.1 mrg 3165 1.1 mrg #define _mm512_srli_epi16(A, B) \ 3166 1.1 mrg ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ 3167 1.3 mrg (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)-1)) 3168 1.1 mrg 3169 1.1 mrg #define _mm512_mask_srli_epi16(W, U, A, B) \ 3170 1.1 mrg ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ 3171 1.1 mrg (int)(B), (__v32hi)(__m512i)(W), (__mmask32)(U))) 3172 1.1 mrg 3173 1.1 mrg #define _mm512_maskz_srli_epi16(U, A, B) \ 3174 1.1 mrg ((__m512i) __builtin_ia32_psrlwi512_mask ((__v32hi)(__m512i)(A), \ 3175 1.3 mrg (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)(U))) 3176 1.1 mrg 3177 1.1 mrg #define _mm512_slli_epi16(X, C) \ 3178 1.1 mrg ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), (int)(C),\ 3179 1.3 mrg (__v32hi)(__m512i)_mm512_setzero_si512 (), \ 3180 1.1 mrg (__mmask32)-1)) 3181 1.1 mrg 3182 1.1 mrg #define _mm512_mask_slli_epi16(W, U, X, C) \ 3183 1.1 mrg ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), (int)(C),\ 3184 1.1 mrg (__v32hi)(__m512i)(W),\ 3185 1.1 mrg (__mmask32)(U))) 3186 1.1 mrg 3187 1.1 mrg #define _mm512_maskz_slli_epi16(U, X, C) \ 3188 1.1 mrg ((__m512i)__builtin_ia32_psllwi512_mask ((__v32hi)(__m512i)(X), (int)(C),\ 3189 1.3 mrg (__v32hi)(__m512i)_mm512_setzero_si512 (), \ 3190 1.1 mrg (__mmask32)(U))) 3191 1.1 mrg 3192 1.1 mrg #define _mm512_shufflehi_epi16(A, B) \ 3193 1.1 mrg ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ 3194 1.3 mrg (__v32hi)(__m512i) \ 3195 1.3 mrg _mm512_setzero_si512 (), \ 3196 1.1 mrg (__mmask32)-1)) 3197 1.1 mrg 3198 1.1 mrg #define _mm512_mask_shufflehi_epi16(W, U, A, B) \ 3199 1.1 mrg ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ 3200 1.1 mrg (__v32hi)(__m512i)(W), \ 3201 1.1 mrg (__mmask32)(U))) 3202 1.1 mrg 3203 1.1 mrg #define _mm512_maskz_shufflehi_epi16(U, A, B) \ 3204 1.1 mrg ((__m512i) __builtin_ia32_pshufhw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ 3205 1.3 mrg (__v32hi)(__m512i) \ 3206 1.3 mrg _mm512_setzero_si512 (), \ 3207 1.1 mrg (__mmask32)(U))) 3208 1.1 mrg 3209 1.1 mrg #define _mm512_shufflelo_epi16(A, B) \ 3210 1.1 mrg ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ 3211 1.3 mrg (__v32hi)(__m512i) \ 3212 1.3 mrg _mm512_setzero_si512 (), \ 3213 1.1 mrg (__mmask32)-1)) 3214 1.1 mrg 3215 1.1 mrg #define _mm512_mask_shufflelo_epi16(W, U, A, B) \ 3216 1.1 mrg ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ 3217 1.1 mrg (__v32hi)(__m512i)(W), \ 3218 1.1 mrg (__mmask32)(U))) 3219 1.1 mrg 3220 1.1 mrg #define _mm512_maskz_shufflelo_epi16(U, A, B) \ 3221 1.1 mrg ((__m512i) __builtin_ia32_pshuflw512_mask ((__v32hi)(__m512i)(A), (int)(B), \ 3222 1.3 mrg (__v32hi)(__m512i) \ 3223 1.3 mrg _mm512_setzero_si512 (), \ 3224 1.1 mrg (__mmask32)(U))) 3225 1.1 mrg 3226 1.1 mrg #define _mm512_srai_epi16(A, B) \ 3227 1.1 mrg ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ 3228 1.3 mrg (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)-1)) 3229 1.1 mrg 3230 1.1 mrg #define _mm512_mask_srai_epi16(W, U, A, B) \ 3231 1.1 mrg ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ 3232 1.1 mrg (int)(B), (__v32hi)(__m512i)(W), (__mmask32)(U))) 3233 1.1 mrg 3234 1.1 mrg #define _mm512_maskz_srai_epi16(U, A, B) \ 3235 1.1 mrg ((__m512i) __builtin_ia32_psrawi512_mask ((__v32hi)(__m512i)(A), \ 3236 1.3 mrg (int)(B), (__v32hi)_mm512_setzero_si512 (), (__mmask32)(U))) 3237 1.1 mrg 3238 1.1 mrg #define _mm512_mask_blend_epi16(__U, __A, __W) \ 3239 1.1 mrg ((__m512i) __builtin_ia32_blendmw_512_mask ((__v32hi) (__A), \ 3240 1.1 mrg (__v32hi) (__W), \ 3241 1.1 mrg (__mmask32) (__U))) 3242 1.1 mrg 3243 1.1 mrg #define _mm512_mask_blend_epi8(__U, __A, __W) \ 3244 1.1 mrg ((__m512i) __builtin_ia32_blendmb_512_mask ((__v64qi) (__A), \ 3245 1.1 mrg (__v64qi) (__W), \ 3246 1.1 mrg (__mmask64) (__U))) 3247 1.1 mrg 3248 1.1 mrg #define _mm512_cmp_epi16_mask(X, Y, P) \ 3249 1.1 mrg ((__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi)(__m512i)(X), \ 3250 1.1 mrg (__v32hi)(__m512i)(Y), (int)(P),\ 3251 1.1 mrg (__mmask32)(-1))) 3252 1.1 mrg 3253 1.1 mrg #define _mm512_cmp_epi8_mask(X, Y, P) \ 3254 1.1 mrg ((__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi)(__m512i)(X), \ 3255 1.1 mrg (__v64qi)(__m512i)(Y), (int)(P),\ 3256 1.1 mrg (__mmask64)(-1))) 3257 1.1 mrg 3258 1.1 mrg #define _mm512_cmp_epu16_mask(X, Y, P) \ 3259 1.1 mrg ((__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi)(__m512i)(X), \ 3260 1.1 mrg (__v32hi)(__m512i)(Y), (int)(P),\ 3261 1.1 mrg (__mmask32)(-1))) 3262 1.1 mrg 3263 1.1 mrg #define _mm512_cmp_epu8_mask(X, Y, P) \ 3264 1.1 mrg ((__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi)(__m512i)(X), \ 3265 1.1 mrg (__v64qi)(__m512i)(Y), (int)(P),\ 3266 1.1 mrg (__mmask64)(-1))) 3267 1.1 mrg 3268 1.1 mrg #define _mm512_mask_cmp_epi16_mask(M, X, Y, P) \ 3269 1.1 mrg ((__mmask32) __builtin_ia32_cmpw512_mask ((__v32hi)(__m512i)(X), \ 3270 1.1 mrg (__v32hi)(__m512i)(Y), (int)(P),\ 3271 1.1 mrg (__mmask32)(M))) 3272 1.1 mrg 3273 1.1 mrg #define _mm512_mask_cmp_epi8_mask(M, X, Y, P) \ 3274 1.1 mrg ((__mmask64) __builtin_ia32_cmpb512_mask ((__v64qi)(__m512i)(X), \ 3275 1.1 mrg (__v64qi)(__m512i)(Y), (int)(P),\ 3276 1.1 mrg (__mmask64)(M))) 3277 1.1 mrg 3278 1.1 mrg #define _mm512_mask_cmp_epu16_mask(M, X, Y, P) \ 3279 1.1 mrg ((__mmask32) __builtin_ia32_ucmpw512_mask ((__v32hi)(__m512i)(X), \ 3280 1.1 mrg (__v32hi)(__m512i)(Y), (int)(P),\ 3281 1.1 mrg (__mmask32)(M))) 3282 1.1 mrg 3283 1.1 mrg #define _mm512_mask_cmp_epu8_mask(M, X, Y, P) \ 3284 1.1 mrg ((__mmask64) __builtin_ia32_ucmpb512_mask ((__v64qi)(__m512i)(X), \ 3285 1.1 mrg (__v64qi)(__m512i)(Y), (int)(P),\ 3286 1.1 mrg (__mmask64)(M))) 3287 1.1 mrg 3288 1.1 mrg #define _mm512_bslli_epi128(A, N) \ 3289 1.1 mrg ((__m512i)__builtin_ia32_pslldq512 ((__m512i)(A), (int)(N) * 8)) 3290 1.1 mrg 3291 1.1 mrg #define _mm512_bsrli_epi128(A, N) \ 3292 1.1 mrg ((__m512i)__builtin_ia32_psrldq512 ((__m512i)(A), (int)(N) * 8)) 3293 1.1 mrg 3294 1.1 mrg #endif 3295 1.1 mrg 3296 1.1 mrg #ifdef __DISABLE_AVX512BW__ 3297 1.1 mrg #undef __DISABLE_AVX512BW__ 3298 1.1 mrg #pragma GCC pop_options 3299 1.1 mrg #endif /* __DISABLE_AVX512BW__ */ 3300 1.1 mrg 3301 1.1 mrg #endif /* _AVX512BWINTRIN_H_INCLUDED */ 3302
Indexes created Wed Mar 25 00:23:37 UTC 2026