avx512dqintrin.h revision 1.4
1 1.4 mrg /* Copyright (C) 2014-2018 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 <avx512dqintrin.h> directly; include <immintrin.h> instead." 26 1.1 mrg #endif 27 1.1 mrg 28 1.1 mrg #ifndef _AVX512DQINTRIN_H_INCLUDED 29 1.1 mrg #define _AVX512DQINTRIN_H_INCLUDED 30 1.1 mrg 31 1.1 mrg #ifndef __AVX512DQ__ 32 1.1 mrg #pragma GCC push_options 33 1.1 mrg #pragma GCC target("avx512dq") 34 1.1 mrg #define __DISABLE_AVX512DQ__ 35 1.1 mrg #endif /* __AVX512DQ__ */ 36 1.1 mrg 37 1.3 mrg extern __inline unsigned char 38 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 39 1.3 mrg _ktest_mask8_u8 (__mmask8 __A, __mmask8 __B, unsigned char *__CF) 40 1.3 mrg { 41 1.3 mrg *__CF = (unsigned char) __builtin_ia32_ktestcqi (__A, __B); 42 1.3 mrg return (unsigned char) __builtin_ia32_ktestzqi (__A, __B); 43 1.3 mrg } 44 1.3 mrg 45 1.3 mrg extern __inline unsigned char 46 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 47 1.3 mrg _ktestz_mask8_u8 (__mmask8 __A, __mmask8 __B) 48 1.3 mrg { 49 1.3 mrg return (unsigned char) __builtin_ia32_ktestzqi (__A, __B); 50 1.3 mrg } 51 1.3 mrg 52 1.3 mrg extern __inline unsigned char 53 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 54 1.3 mrg _ktestc_mask8_u8 (__mmask8 __A, __mmask8 __B) 55 1.3 mrg { 56 1.3 mrg return (unsigned char) __builtin_ia32_ktestcqi (__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 _ktest_mask16_u8 (__mmask16 __A, __mmask16 __B, unsigned char *__CF) 62 1.3 mrg { 63 1.3 mrg *__CF = (unsigned char) __builtin_ia32_ktestchi (__A, __B); 64 1.3 mrg return (unsigned char) __builtin_ia32_ktestzhi (__A, __B); 65 1.3 mrg } 66 1.3 mrg 67 1.3 mrg extern __inline unsigned char 68 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 69 1.3 mrg _ktestz_mask16_u8 (__mmask16 __A, __mmask16 __B) 70 1.3 mrg { 71 1.3 mrg return (unsigned char) __builtin_ia32_ktestzhi (__A, __B); 72 1.3 mrg } 73 1.3 mrg 74 1.3 mrg extern __inline unsigned char 75 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 76 1.3 mrg _ktestc_mask16_u8 (__mmask16 __A, __mmask16 __B) 77 1.3 mrg { 78 1.3 mrg return (unsigned char) __builtin_ia32_ktestchi (__A, __B); 79 1.3 mrg } 80 1.3 mrg 81 1.3 mrg extern __inline unsigned char 82 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 83 1.3 mrg _kortest_mask8_u8 (__mmask8 __A, __mmask8 __B, unsigned char *__CF) 84 1.3 mrg { 85 1.3 mrg *__CF = (unsigned char) __builtin_ia32_kortestcqi (__A, __B); 86 1.3 mrg return (unsigned char) __builtin_ia32_kortestzqi (__A, __B); 87 1.3 mrg } 88 1.3 mrg 89 1.3 mrg extern __inline unsigned char 90 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 91 1.3 mrg _kortestz_mask8_u8 (__mmask8 __A, __mmask8 __B) 92 1.3 mrg { 93 1.3 mrg return (unsigned char) __builtin_ia32_kortestzqi (__A, __B); 94 1.3 mrg } 95 1.3 mrg 96 1.3 mrg extern __inline unsigned char 97 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 98 1.3 mrg _kortestc_mask8_u8 (__mmask8 __A, __mmask8 __B) 99 1.3 mrg { 100 1.3 mrg return (unsigned char) __builtin_ia32_kortestcqi (__A, __B); 101 1.3 mrg } 102 1.3 mrg 103 1.3 mrg extern __inline __mmask8 104 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 105 1.3 mrg _kadd_mask8 (__mmask8 __A, __mmask8 __B) 106 1.3 mrg { 107 1.3 mrg return (__mmask8) __builtin_ia32_kaddqi ((__mmask8) __A, (__mmask8) __B); 108 1.3 mrg } 109 1.3 mrg 110 1.3 mrg extern __inline __mmask16 111 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 112 1.3 mrg _kadd_mask16 (__mmask16 __A, __mmask16 __B) 113 1.3 mrg { 114 1.3 mrg return (__mmask16) __builtin_ia32_kaddhi ((__mmask16) __A, (__mmask16) __B); 115 1.3 mrg } 116 1.3 mrg 117 1.3 mrg extern __inline unsigned int 118 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 119 1.3 mrg _cvtmask8_u32 (__mmask8 __A) 120 1.3 mrg { 121 1.3 mrg return (unsigned int) __builtin_ia32_kmovb ((__mmask8 ) __A); 122 1.3 mrg } 123 1.3 mrg 124 1.3 mrg extern __inline __mmask8 125 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 126 1.3 mrg _cvtu32_mask8 (unsigned int __A) 127 1.3 mrg { 128 1.3 mrg return (__mmask8) __builtin_ia32_kmovb ((__mmask8) __A); 129 1.3 mrg } 130 1.3 mrg 131 1.3 mrg extern __inline __mmask8 132 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 133 1.3 mrg _load_mask8 (__mmask8 *__A) 134 1.3 mrg { 135 1.3 mrg return (__mmask8) __builtin_ia32_kmovb (*(__mmask8 *) __A); 136 1.3 mrg } 137 1.3 mrg 138 1.3 mrg extern __inline void 139 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 140 1.3 mrg _store_mask8 (__mmask8 *__A, __mmask8 __B) 141 1.3 mrg { 142 1.3 mrg *(__mmask8 *) __A = __builtin_ia32_kmovb (__B); 143 1.3 mrg } 144 1.3 mrg 145 1.3 mrg extern __inline __mmask8 146 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 147 1.3 mrg _knot_mask8 (__mmask8 __A) 148 1.3 mrg { 149 1.3 mrg return (__mmask8) __builtin_ia32_knotqi ((__mmask8) __A); 150 1.3 mrg } 151 1.3 mrg 152 1.3 mrg extern __inline __mmask8 153 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 154 1.3 mrg _kor_mask8 (__mmask8 __A, __mmask8 __B) 155 1.3 mrg { 156 1.3 mrg return (__mmask8) __builtin_ia32_korqi ((__mmask8) __A, (__mmask8) __B); 157 1.3 mrg } 158 1.3 mrg 159 1.3 mrg extern __inline __mmask8 160 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 161 1.3 mrg _kxnor_mask8 (__mmask8 __A, __mmask8 __B) 162 1.3 mrg { 163 1.3 mrg return (__mmask8) __builtin_ia32_kxnorqi ((__mmask8) __A, (__mmask8) __B); 164 1.3 mrg } 165 1.3 mrg 166 1.3 mrg extern __inline __mmask8 167 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 168 1.3 mrg _kxor_mask8 (__mmask8 __A, __mmask8 __B) 169 1.3 mrg { 170 1.3 mrg return (__mmask8) __builtin_ia32_kxorqi ((__mmask8) __A, (__mmask8) __B); 171 1.3 mrg } 172 1.3 mrg 173 1.3 mrg extern __inline __mmask8 174 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 175 1.3 mrg _kand_mask8 (__mmask8 __A, __mmask8 __B) 176 1.3 mrg { 177 1.3 mrg return (__mmask8) __builtin_ia32_kandqi ((__mmask8) __A, (__mmask8) __B); 178 1.3 mrg } 179 1.3 mrg 180 1.3 mrg extern __inline __mmask8 181 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 182 1.3 mrg _kandn_mask8 (__mmask8 __A, __mmask8 __B) 183 1.3 mrg { 184 1.3 mrg return (__mmask8) __builtin_ia32_kandnqi ((__mmask8) __A, (__mmask8) __B); 185 1.3 mrg } 186 1.3 mrg 187 1.1 mrg extern __inline __m512d 188 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 189 1.1 mrg _mm512_broadcast_f64x2 (__m128d __A) 190 1.1 mrg { 191 1.3 mrg return (__m512d) 192 1.3 mrg __builtin_ia32_broadcastf64x2_512_mask ((__v2df) __A, 193 1.3 mrg _mm512_undefined_pd (), 194 1.3 mrg (__mmask8) -1); 195 1.1 mrg } 196 1.1 mrg 197 1.1 mrg extern __inline __m512d 198 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 199 1.1 mrg _mm512_mask_broadcast_f64x2 (__m512d __O, __mmask8 __M, __m128d __A) 200 1.1 mrg { 201 1.1 mrg return (__m512d) __builtin_ia32_broadcastf64x2_512_mask ((__v2df) 202 1.1 mrg __A, 203 1.1 mrg (__v8df) 204 1.1 mrg __O, __M); 205 1.1 mrg } 206 1.1 mrg 207 1.1 mrg extern __inline __m512d 208 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 209 1.1 mrg _mm512_maskz_broadcast_f64x2 (__mmask8 __M, __m128d __A) 210 1.1 mrg { 211 1.1 mrg return (__m512d) __builtin_ia32_broadcastf64x2_512_mask ((__v2df) 212 1.1 mrg __A, 213 1.1 mrg (__v8df) 214 1.1 mrg _mm512_setzero_ps (), 215 1.1 mrg __M); 216 1.1 mrg } 217 1.1 mrg 218 1.1 mrg extern __inline __m512i 219 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 220 1.1 mrg _mm512_broadcast_i64x2 (__m128i __A) 221 1.1 mrg { 222 1.3 mrg return (__m512i) 223 1.3 mrg __builtin_ia32_broadcasti64x2_512_mask ((__v2di) __A, 224 1.3 mrg _mm512_undefined_epi32 (), 225 1.3 mrg (__mmask8) -1); 226 1.1 mrg } 227 1.1 mrg 228 1.1 mrg extern __inline __m512i 229 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 230 1.1 mrg _mm512_mask_broadcast_i64x2 (__m512i __O, __mmask8 __M, __m128i __A) 231 1.1 mrg { 232 1.1 mrg return (__m512i) __builtin_ia32_broadcasti64x2_512_mask ((__v2di) 233 1.1 mrg __A, 234 1.1 mrg (__v8di) 235 1.1 mrg __O, __M); 236 1.1 mrg } 237 1.1 mrg 238 1.1 mrg extern __inline __m512i 239 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 240 1.1 mrg _mm512_maskz_broadcast_i64x2 (__mmask8 __M, __m128i __A) 241 1.1 mrg { 242 1.1 mrg return (__m512i) __builtin_ia32_broadcasti64x2_512_mask ((__v2di) 243 1.1 mrg __A, 244 1.1 mrg (__v8di) 245 1.1 mrg _mm512_setzero_si512 (), 246 1.1 mrg __M); 247 1.1 mrg } 248 1.1 mrg 249 1.1 mrg extern __inline __m512 250 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 251 1.1 mrg _mm512_broadcast_f32x2 (__m128 __A) 252 1.1 mrg { 253 1.3 mrg return (__m512) 254 1.3 mrg __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, 255 1.3 mrg (__v16sf)_mm512_undefined_ps (), 256 1.3 mrg (__mmask16) -1); 257 1.1 mrg } 258 1.1 mrg 259 1.1 mrg extern __inline __m512 260 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 261 1.1 mrg _mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A) 262 1.1 mrg { 263 1.1 mrg return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, 264 1.1 mrg (__v16sf) 265 1.1 mrg __O, __M); 266 1.1 mrg } 267 1.1 mrg 268 1.1 mrg extern __inline __m512 269 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 270 1.1 mrg _mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A) 271 1.1 mrg { 272 1.1 mrg return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, 273 1.1 mrg (__v16sf) 274 1.1 mrg _mm512_setzero_ps (), 275 1.1 mrg __M); 276 1.1 mrg } 277 1.1 mrg 278 1.1 mrg extern __inline __m512i 279 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 280 1.1 mrg _mm512_broadcast_i32x2 (__m128i __A) 281 1.1 mrg { 282 1.3 mrg return (__m512i) 283 1.3 mrg __builtin_ia32_broadcasti32x2_512_mask ((__v4si) __A, 284 1.3 mrg (__v16si) 285 1.3 mrg _mm512_undefined_epi32 (), 286 1.3 mrg (__mmask16) -1); 287 1.1 mrg } 288 1.1 mrg 289 1.1 mrg extern __inline __m512i 290 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 291 1.1 mrg _mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A) 292 1.1 mrg { 293 1.1 mrg return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) 294 1.1 mrg __A, 295 1.1 mrg (__v16si) 296 1.1 mrg __O, __M); 297 1.1 mrg } 298 1.1 mrg 299 1.1 mrg extern __inline __m512i 300 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 301 1.1 mrg _mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A) 302 1.1 mrg { 303 1.1 mrg return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) 304 1.1 mrg __A, 305 1.1 mrg (__v16si) 306 1.1 mrg _mm512_setzero_si512 (), 307 1.1 mrg __M); 308 1.1 mrg } 309 1.1 mrg 310 1.1 mrg extern __inline __m512 311 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 312 1.1 mrg _mm512_broadcast_f32x8 (__m256 __A) 313 1.1 mrg { 314 1.3 mrg return (__m512) 315 1.3 mrg __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, 316 1.3 mrg _mm512_undefined_ps (), 317 1.3 mrg (__mmask16) -1); 318 1.1 mrg } 319 1.1 mrg 320 1.1 mrg extern __inline __m512 321 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 322 1.1 mrg _mm512_mask_broadcast_f32x8 (__m512 __O, __mmask16 __M, __m256 __A) 323 1.1 mrg { 324 1.1 mrg return (__m512) __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, 325 1.1 mrg (__v16sf)__O, 326 1.1 mrg __M); 327 1.1 mrg } 328 1.1 mrg 329 1.1 mrg extern __inline __m512 330 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 331 1.1 mrg _mm512_maskz_broadcast_f32x8 (__mmask16 __M, __m256 __A) 332 1.1 mrg { 333 1.1 mrg return (__m512) __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, 334 1.1 mrg (__v16sf) 335 1.1 mrg _mm512_setzero_ps (), 336 1.1 mrg __M); 337 1.1 mrg } 338 1.1 mrg 339 1.1 mrg extern __inline __m512i 340 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 341 1.1 mrg _mm512_broadcast_i32x8 (__m256i __A) 342 1.1 mrg { 343 1.3 mrg return (__m512i) 344 1.3 mrg __builtin_ia32_broadcasti32x8_512_mask ((__v8si) __A, 345 1.3 mrg (__v16si) 346 1.3 mrg _mm512_undefined_epi32 (), 347 1.3 mrg (__mmask16) -1); 348 1.1 mrg } 349 1.1 mrg 350 1.1 mrg extern __inline __m512i 351 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 352 1.1 mrg _mm512_mask_broadcast_i32x8 (__m512i __O, __mmask16 __M, __m256i __A) 353 1.1 mrg { 354 1.1 mrg return (__m512i) __builtin_ia32_broadcasti32x8_512_mask ((__v8si) 355 1.1 mrg __A, 356 1.1 mrg (__v16si)__O, 357 1.1 mrg __M); 358 1.1 mrg } 359 1.1 mrg 360 1.1 mrg extern __inline __m512i 361 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 362 1.1 mrg _mm512_maskz_broadcast_i32x8 (__mmask16 __M, __m256i __A) 363 1.1 mrg { 364 1.1 mrg return (__m512i) __builtin_ia32_broadcasti32x8_512_mask ((__v8si) 365 1.1 mrg __A, 366 1.1 mrg (__v16si) 367 1.1 mrg _mm512_setzero_si512 (), 368 1.1 mrg __M); 369 1.1 mrg } 370 1.1 mrg 371 1.1 mrg extern __inline __m512i 372 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 373 1.1 mrg _mm512_mullo_epi64 (__m512i __A, __m512i __B) 374 1.1 mrg { 375 1.1 mrg return (__m512i) ((__v8du) __A * (__v8du) __B); 376 1.1 mrg } 377 1.1 mrg 378 1.1 mrg extern __inline __m512i 379 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 380 1.1 mrg _mm512_mask_mullo_epi64 (__m512i __W, __mmask8 __U, __m512i __A, 381 1.1 mrg __m512i __B) 382 1.1 mrg { 383 1.1 mrg return (__m512i) __builtin_ia32_pmullq512_mask ((__v8di) __A, 384 1.1 mrg (__v8di) __B, 385 1.1 mrg (__v8di) __W, 386 1.1 mrg (__mmask8) __U); 387 1.1 mrg } 388 1.1 mrg 389 1.1 mrg extern __inline __m512i 390 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 391 1.1 mrg _mm512_maskz_mullo_epi64 (__mmask8 __U, __m512i __A, __m512i __B) 392 1.1 mrg { 393 1.1 mrg return (__m512i) __builtin_ia32_pmullq512_mask ((__v8di) __A, 394 1.1 mrg (__v8di) __B, 395 1.1 mrg (__v8di) 396 1.1 mrg _mm512_setzero_si512 (), 397 1.1 mrg (__mmask8) __U); 398 1.1 mrg } 399 1.1 mrg 400 1.1 mrg extern __inline __m512d 401 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 402 1.1 mrg _mm512_xor_pd (__m512d __A, __m512d __B) 403 1.1 mrg { 404 1.1 mrg return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, 405 1.1 mrg (__v8df) __B, 406 1.1 mrg (__v8df) 407 1.1 mrg _mm512_setzero_pd (), 408 1.1 mrg (__mmask8) -1); 409 1.1 mrg } 410 1.1 mrg 411 1.1 mrg extern __inline __m512d 412 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 413 1.1 mrg _mm512_mask_xor_pd (__m512d __W, __mmask8 __U, __m512d __A, 414 1.1 mrg __m512d __B) 415 1.1 mrg { 416 1.1 mrg return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, 417 1.1 mrg (__v8df) __B, 418 1.1 mrg (__v8df) __W, 419 1.1 mrg (__mmask8) __U); 420 1.1 mrg } 421 1.1 mrg 422 1.1 mrg extern __inline __m512d 423 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 424 1.1 mrg _mm512_maskz_xor_pd (__mmask8 __U, __m512d __A, __m512d __B) 425 1.1 mrg { 426 1.1 mrg return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, 427 1.1 mrg (__v8df) __B, 428 1.1 mrg (__v8df) 429 1.1 mrg _mm512_setzero_pd (), 430 1.1 mrg (__mmask8) __U); 431 1.1 mrg } 432 1.1 mrg 433 1.1 mrg extern __inline __m512 434 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 435 1.1 mrg _mm512_xor_ps (__m512 __A, __m512 __B) 436 1.1 mrg { 437 1.1 mrg return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, 438 1.1 mrg (__v16sf) __B, 439 1.1 mrg (__v16sf) 440 1.1 mrg _mm512_setzero_ps (), 441 1.1 mrg (__mmask16) -1); 442 1.1 mrg } 443 1.1 mrg 444 1.1 mrg extern __inline __m512 445 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 446 1.1 mrg _mm512_mask_xor_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) 447 1.1 mrg { 448 1.1 mrg return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, 449 1.1 mrg (__v16sf) __B, 450 1.1 mrg (__v16sf) __W, 451 1.1 mrg (__mmask16) __U); 452 1.1 mrg } 453 1.1 mrg 454 1.1 mrg extern __inline __m512 455 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 456 1.1 mrg _mm512_maskz_xor_ps (__mmask16 __U, __m512 __A, __m512 __B) 457 1.1 mrg { 458 1.1 mrg return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, 459 1.1 mrg (__v16sf) __B, 460 1.1 mrg (__v16sf) 461 1.1 mrg _mm512_setzero_ps (), 462 1.1 mrg (__mmask16) __U); 463 1.1 mrg } 464 1.1 mrg 465 1.1 mrg extern __inline __m512d 466 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 467 1.1 mrg _mm512_or_pd (__m512d __A, __m512d __B) 468 1.1 mrg { 469 1.1 mrg return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, 470 1.1 mrg (__v8df) __B, 471 1.1 mrg (__v8df) 472 1.1 mrg _mm512_setzero_pd (), 473 1.1 mrg (__mmask8) -1); 474 1.1 mrg } 475 1.1 mrg 476 1.1 mrg extern __inline __m512d 477 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 478 1.1 mrg _mm512_mask_or_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) 479 1.1 mrg { 480 1.1 mrg return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, 481 1.1 mrg (__v8df) __B, 482 1.1 mrg (__v8df) __W, 483 1.1 mrg (__mmask8) __U); 484 1.1 mrg } 485 1.1 mrg 486 1.1 mrg extern __inline __m512d 487 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 488 1.1 mrg _mm512_maskz_or_pd (__mmask8 __U, __m512d __A, __m512d __B) 489 1.1 mrg { 490 1.1 mrg return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, 491 1.1 mrg (__v8df) __B, 492 1.1 mrg (__v8df) 493 1.1 mrg _mm512_setzero_pd (), 494 1.1 mrg (__mmask8) __U); 495 1.1 mrg } 496 1.1 mrg 497 1.1 mrg extern __inline __m512 498 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 499 1.1 mrg _mm512_or_ps (__m512 __A, __m512 __B) 500 1.1 mrg { 501 1.1 mrg return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, 502 1.1 mrg (__v16sf) __B, 503 1.1 mrg (__v16sf) 504 1.1 mrg _mm512_setzero_ps (), 505 1.1 mrg (__mmask16) -1); 506 1.1 mrg } 507 1.1 mrg 508 1.1 mrg extern __inline __m512 509 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 510 1.1 mrg _mm512_mask_or_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) 511 1.1 mrg { 512 1.1 mrg return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, 513 1.1 mrg (__v16sf) __B, 514 1.1 mrg (__v16sf) __W, 515 1.1 mrg (__mmask16) __U); 516 1.1 mrg } 517 1.1 mrg 518 1.1 mrg extern __inline __m512 519 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 520 1.1 mrg _mm512_maskz_or_ps (__mmask16 __U, __m512 __A, __m512 __B) 521 1.1 mrg { 522 1.1 mrg return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, 523 1.1 mrg (__v16sf) __B, 524 1.1 mrg (__v16sf) 525 1.1 mrg _mm512_setzero_ps (), 526 1.1 mrg (__mmask16) __U); 527 1.1 mrg } 528 1.1 mrg 529 1.1 mrg extern __inline __m512d 530 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 531 1.1 mrg _mm512_and_pd (__m512d __A, __m512d __B) 532 1.1 mrg { 533 1.1 mrg return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, 534 1.1 mrg (__v8df) __B, 535 1.1 mrg (__v8df) 536 1.1 mrg _mm512_setzero_pd (), 537 1.1 mrg (__mmask8) -1); 538 1.1 mrg } 539 1.1 mrg 540 1.1 mrg extern __inline __m512d 541 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 542 1.1 mrg _mm512_mask_and_pd (__m512d __W, __mmask8 __U, __m512d __A, 543 1.1 mrg __m512d __B) 544 1.1 mrg { 545 1.1 mrg return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, 546 1.1 mrg (__v8df) __B, 547 1.1 mrg (__v8df) __W, 548 1.1 mrg (__mmask8) __U); 549 1.1 mrg } 550 1.1 mrg 551 1.1 mrg extern __inline __m512d 552 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 553 1.1 mrg _mm512_maskz_and_pd (__mmask8 __U, __m512d __A, __m512d __B) 554 1.1 mrg { 555 1.1 mrg return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, 556 1.1 mrg (__v8df) __B, 557 1.1 mrg (__v8df) 558 1.1 mrg _mm512_setzero_pd (), 559 1.1 mrg (__mmask8) __U); 560 1.1 mrg } 561 1.1 mrg 562 1.1 mrg extern __inline __m512 563 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 564 1.1 mrg _mm512_and_ps (__m512 __A, __m512 __B) 565 1.1 mrg { 566 1.1 mrg return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, 567 1.1 mrg (__v16sf) __B, 568 1.1 mrg (__v16sf) 569 1.1 mrg _mm512_setzero_ps (), 570 1.1 mrg (__mmask16) -1); 571 1.1 mrg } 572 1.1 mrg 573 1.1 mrg extern __inline __m512 574 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 575 1.1 mrg _mm512_mask_and_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) 576 1.1 mrg { 577 1.1 mrg return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, 578 1.1 mrg (__v16sf) __B, 579 1.1 mrg (__v16sf) __W, 580 1.1 mrg (__mmask16) __U); 581 1.1 mrg } 582 1.1 mrg 583 1.1 mrg extern __inline __m512 584 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 585 1.1 mrg _mm512_maskz_and_ps (__mmask16 __U, __m512 __A, __m512 __B) 586 1.1 mrg { 587 1.1 mrg return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, 588 1.1 mrg (__v16sf) __B, 589 1.1 mrg (__v16sf) 590 1.1 mrg _mm512_setzero_ps (), 591 1.1 mrg (__mmask16) __U); 592 1.1 mrg } 593 1.1 mrg 594 1.1 mrg extern __inline __m512d 595 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 596 1.1 mrg _mm512_andnot_pd (__m512d __A, __m512d __B) 597 1.1 mrg { 598 1.1 mrg return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, 599 1.1 mrg (__v8df) __B, 600 1.1 mrg (__v8df) 601 1.1 mrg _mm512_setzero_pd (), 602 1.1 mrg (__mmask8) -1); 603 1.1 mrg } 604 1.1 mrg 605 1.1 mrg extern __inline __m512d 606 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 607 1.1 mrg _mm512_mask_andnot_pd (__m512d __W, __mmask8 __U, __m512d __A, 608 1.1 mrg __m512d __B) 609 1.1 mrg { 610 1.1 mrg return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, 611 1.1 mrg (__v8df) __B, 612 1.1 mrg (__v8df) __W, 613 1.1 mrg (__mmask8) __U); 614 1.1 mrg } 615 1.1 mrg 616 1.1 mrg extern __inline __m512d 617 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 618 1.1 mrg _mm512_maskz_andnot_pd (__mmask8 __U, __m512d __A, __m512d __B) 619 1.1 mrg { 620 1.1 mrg return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, 621 1.1 mrg (__v8df) __B, 622 1.1 mrg (__v8df) 623 1.1 mrg _mm512_setzero_pd (), 624 1.1 mrg (__mmask8) __U); 625 1.1 mrg } 626 1.1 mrg 627 1.1 mrg extern __inline __m512 628 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 629 1.1 mrg _mm512_andnot_ps (__m512 __A, __m512 __B) 630 1.1 mrg { 631 1.1 mrg return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, 632 1.1 mrg (__v16sf) __B, 633 1.1 mrg (__v16sf) 634 1.1 mrg _mm512_setzero_ps (), 635 1.1 mrg (__mmask16) -1); 636 1.1 mrg } 637 1.1 mrg 638 1.1 mrg extern __inline __m512 639 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 640 1.1 mrg _mm512_mask_andnot_ps (__m512 __W, __mmask16 __U, __m512 __A, 641 1.1 mrg __m512 __B) 642 1.1 mrg { 643 1.1 mrg return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, 644 1.1 mrg (__v16sf) __B, 645 1.1 mrg (__v16sf) __W, 646 1.1 mrg (__mmask16) __U); 647 1.1 mrg } 648 1.1 mrg 649 1.1 mrg extern __inline __m512 650 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 651 1.1 mrg _mm512_maskz_andnot_ps (__mmask16 __U, __m512 __A, __m512 __B) 652 1.1 mrg { 653 1.1 mrg return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, 654 1.1 mrg (__v16sf) __B, 655 1.1 mrg (__v16sf) 656 1.1 mrg _mm512_setzero_ps (), 657 1.1 mrg (__mmask16) __U); 658 1.1 mrg } 659 1.1 mrg 660 1.1 mrg extern __inline __mmask16 661 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 662 1.1 mrg _mm512_movepi32_mask (__m512i __A) 663 1.1 mrg { 664 1.1 mrg return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A); 665 1.1 mrg } 666 1.1 mrg 667 1.1 mrg extern __inline __mmask8 668 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 669 1.1 mrg _mm512_movepi64_mask (__m512i __A) 670 1.1 mrg { 671 1.1 mrg return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A); 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_movm_epi32 (__mmask16 __A) 677 1.1 mrg { 678 1.1 mrg return (__m512i) __builtin_ia32_cvtmask2d512 (__A); 679 1.1 mrg } 680 1.1 mrg 681 1.1 mrg extern __inline __m512i 682 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 683 1.1 mrg _mm512_movm_epi64 (__mmask8 __A) 684 1.1 mrg { 685 1.1 mrg return (__m512i) __builtin_ia32_cvtmask2q512 (__A); 686 1.1 mrg } 687 1.1 mrg 688 1.1 mrg extern __inline __m512i 689 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 690 1.1 mrg _mm512_cvttpd_epi64 (__m512d __A) 691 1.1 mrg { 692 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 693 1.1 mrg (__v8di) 694 1.1 mrg _mm512_setzero_si512 (), 695 1.1 mrg (__mmask8) -1, 696 1.1 mrg _MM_FROUND_CUR_DIRECTION); 697 1.1 mrg } 698 1.1 mrg 699 1.1 mrg extern __inline __m512i 700 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 701 1.1 mrg _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) 702 1.1 mrg { 703 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 704 1.1 mrg (__v8di) __W, 705 1.1 mrg (__mmask8) __U, 706 1.1 mrg _MM_FROUND_CUR_DIRECTION); 707 1.1 mrg } 708 1.1 mrg 709 1.1 mrg extern __inline __m512i 710 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 711 1.1 mrg _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) 712 1.1 mrg { 713 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 714 1.1 mrg (__v8di) 715 1.1 mrg _mm512_setzero_si512 (), 716 1.1 mrg (__mmask8) __U, 717 1.1 mrg _MM_FROUND_CUR_DIRECTION); 718 1.1 mrg } 719 1.1 mrg 720 1.1 mrg extern __inline __m512i 721 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 722 1.1 mrg _mm512_cvttpd_epu64 (__m512d __A) 723 1.1 mrg { 724 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 725 1.1 mrg (__v8di) 726 1.1 mrg _mm512_setzero_si512 (), 727 1.1 mrg (__mmask8) -1, 728 1.1 mrg _MM_FROUND_CUR_DIRECTION); 729 1.1 mrg } 730 1.1 mrg 731 1.1 mrg extern __inline __m512i 732 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 733 1.1 mrg _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) 734 1.1 mrg { 735 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 736 1.1 mrg (__v8di) __W, 737 1.1 mrg (__mmask8) __U, 738 1.1 mrg _MM_FROUND_CUR_DIRECTION); 739 1.1 mrg } 740 1.1 mrg 741 1.1 mrg extern __inline __m512i 742 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 743 1.1 mrg _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) 744 1.1 mrg { 745 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 746 1.1 mrg (__v8di) 747 1.1 mrg _mm512_setzero_si512 (), 748 1.1 mrg (__mmask8) __U, 749 1.1 mrg _MM_FROUND_CUR_DIRECTION); 750 1.1 mrg } 751 1.1 mrg 752 1.1 mrg extern __inline __m512i 753 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 754 1.1 mrg _mm512_cvttps_epi64 (__m256 __A) 755 1.1 mrg { 756 1.1 mrg return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 757 1.1 mrg (__v8di) 758 1.1 mrg _mm512_setzero_si512 (), 759 1.1 mrg (__mmask8) -1, 760 1.1 mrg _MM_FROUND_CUR_DIRECTION); 761 1.1 mrg } 762 1.1 mrg 763 1.1 mrg extern __inline __m512i 764 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 765 1.1 mrg _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) 766 1.1 mrg { 767 1.1 mrg return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 768 1.1 mrg (__v8di) __W, 769 1.1 mrg (__mmask8) __U, 770 1.1 mrg _MM_FROUND_CUR_DIRECTION); 771 1.1 mrg } 772 1.1 mrg 773 1.1 mrg extern __inline __m512i 774 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 775 1.1 mrg _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) 776 1.1 mrg { 777 1.1 mrg return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 778 1.1 mrg (__v8di) 779 1.1 mrg _mm512_setzero_si512 (), 780 1.1 mrg (__mmask8) __U, 781 1.1 mrg _MM_FROUND_CUR_DIRECTION); 782 1.1 mrg } 783 1.1 mrg 784 1.1 mrg extern __inline __m512i 785 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 786 1.1 mrg _mm512_cvttps_epu64 (__m256 __A) 787 1.1 mrg { 788 1.1 mrg return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 789 1.1 mrg (__v8di) 790 1.1 mrg _mm512_setzero_si512 (), 791 1.1 mrg (__mmask8) -1, 792 1.1 mrg _MM_FROUND_CUR_DIRECTION); 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_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) 798 1.1 mrg { 799 1.1 mrg return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 800 1.1 mrg (__v8di) __W, 801 1.1 mrg (__mmask8) __U, 802 1.1 mrg _MM_FROUND_CUR_DIRECTION); 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_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) 808 1.1 mrg { 809 1.1 mrg return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 810 1.1 mrg (__v8di) 811 1.1 mrg _mm512_setzero_si512 (), 812 1.1 mrg (__mmask8) __U, 813 1.1 mrg _MM_FROUND_CUR_DIRECTION); 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_cvtpd_epi64 (__m512d __A) 819 1.1 mrg { 820 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 821 1.1 mrg (__v8di) 822 1.1 mrg _mm512_setzero_si512 (), 823 1.1 mrg (__mmask8) -1, 824 1.1 mrg _MM_FROUND_CUR_DIRECTION); 825 1.1 mrg } 826 1.1 mrg 827 1.1 mrg extern __inline __m512i 828 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 829 1.1 mrg _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) 830 1.1 mrg { 831 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 832 1.1 mrg (__v8di) __W, 833 1.1 mrg (__mmask8) __U, 834 1.1 mrg _MM_FROUND_CUR_DIRECTION); 835 1.1 mrg } 836 1.1 mrg 837 1.1 mrg extern __inline __m512i 838 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 839 1.1 mrg _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) 840 1.1 mrg { 841 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 842 1.1 mrg (__v8di) 843 1.1 mrg _mm512_setzero_si512 (), 844 1.1 mrg (__mmask8) __U, 845 1.1 mrg _MM_FROUND_CUR_DIRECTION); 846 1.1 mrg } 847 1.1 mrg 848 1.1 mrg extern __inline __m512i 849 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 850 1.1 mrg _mm512_cvtpd_epu64 (__m512d __A) 851 1.1 mrg { 852 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 853 1.1 mrg (__v8di) 854 1.1 mrg _mm512_setzero_si512 (), 855 1.1 mrg (__mmask8) -1, 856 1.1 mrg _MM_FROUND_CUR_DIRECTION); 857 1.1 mrg } 858 1.1 mrg 859 1.1 mrg extern __inline __m512i 860 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 861 1.1 mrg _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) 862 1.1 mrg { 863 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 864 1.1 mrg (__v8di) __W, 865 1.1 mrg (__mmask8) __U, 866 1.1 mrg _MM_FROUND_CUR_DIRECTION); 867 1.1 mrg } 868 1.1 mrg 869 1.1 mrg extern __inline __m512i 870 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 871 1.1 mrg _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) 872 1.1 mrg { 873 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 874 1.1 mrg (__v8di) 875 1.1 mrg _mm512_setzero_si512 (), 876 1.1 mrg (__mmask8) __U, 877 1.1 mrg _MM_FROUND_CUR_DIRECTION); 878 1.1 mrg } 879 1.1 mrg 880 1.1 mrg extern __inline __m512i 881 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 882 1.1 mrg _mm512_cvtps_epi64 (__m256 __A) 883 1.1 mrg { 884 1.1 mrg return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 885 1.1 mrg (__v8di) 886 1.1 mrg _mm512_setzero_si512 (), 887 1.1 mrg (__mmask8) -1, 888 1.1 mrg _MM_FROUND_CUR_DIRECTION); 889 1.1 mrg } 890 1.1 mrg 891 1.1 mrg extern __inline __m512i 892 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 893 1.1 mrg _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) 894 1.1 mrg { 895 1.1 mrg return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 896 1.1 mrg (__v8di) __W, 897 1.1 mrg (__mmask8) __U, 898 1.1 mrg _MM_FROUND_CUR_DIRECTION); 899 1.1 mrg } 900 1.1 mrg 901 1.1 mrg extern __inline __m512i 902 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 903 1.1 mrg _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) 904 1.1 mrg { 905 1.1 mrg return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 906 1.1 mrg (__v8di) 907 1.1 mrg _mm512_setzero_si512 (), 908 1.1 mrg (__mmask8) __U, 909 1.1 mrg _MM_FROUND_CUR_DIRECTION); 910 1.1 mrg } 911 1.1 mrg 912 1.1 mrg extern __inline __m512i 913 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 914 1.1 mrg _mm512_cvtps_epu64 (__m256 __A) 915 1.1 mrg { 916 1.1 mrg return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 917 1.1 mrg (__v8di) 918 1.1 mrg _mm512_setzero_si512 (), 919 1.1 mrg (__mmask8) -1, 920 1.1 mrg _MM_FROUND_CUR_DIRECTION); 921 1.1 mrg } 922 1.1 mrg 923 1.1 mrg extern __inline __m512i 924 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 925 1.1 mrg _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) 926 1.1 mrg { 927 1.1 mrg return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 928 1.1 mrg (__v8di) __W, 929 1.1 mrg (__mmask8) __U, 930 1.1 mrg _MM_FROUND_CUR_DIRECTION); 931 1.1 mrg } 932 1.1 mrg 933 1.1 mrg extern __inline __m512i 934 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 935 1.1 mrg _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) 936 1.1 mrg { 937 1.1 mrg return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 938 1.1 mrg (__v8di) 939 1.1 mrg _mm512_setzero_si512 (), 940 1.1 mrg (__mmask8) __U, 941 1.1 mrg _MM_FROUND_CUR_DIRECTION); 942 1.1 mrg } 943 1.1 mrg 944 1.1 mrg extern __inline __m256 945 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 946 1.1 mrg _mm512_cvtepi64_ps (__m512i __A) 947 1.1 mrg { 948 1.1 mrg return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 949 1.1 mrg (__v8sf) 950 1.1 mrg _mm256_setzero_ps (), 951 1.1 mrg (__mmask8) -1, 952 1.1 mrg _MM_FROUND_CUR_DIRECTION); 953 1.1 mrg } 954 1.1 mrg 955 1.1 mrg extern __inline __m256 956 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 957 1.1 mrg _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) 958 1.1 mrg { 959 1.1 mrg return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 960 1.1 mrg (__v8sf) __W, 961 1.1 mrg (__mmask8) __U, 962 1.1 mrg _MM_FROUND_CUR_DIRECTION); 963 1.1 mrg } 964 1.1 mrg 965 1.1 mrg extern __inline __m256 966 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 967 1.1 mrg _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) 968 1.1 mrg { 969 1.1 mrg return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 970 1.1 mrg (__v8sf) 971 1.1 mrg _mm256_setzero_ps (), 972 1.1 mrg (__mmask8) __U, 973 1.1 mrg _MM_FROUND_CUR_DIRECTION); 974 1.1 mrg } 975 1.1 mrg 976 1.1 mrg extern __inline __m256 977 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 978 1.1 mrg _mm512_cvtepu64_ps (__m512i __A) 979 1.1 mrg { 980 1.1 mrg return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 981 1.1 mrg (__v8sf) 982 1.1 mrg _mm256_setzero_ps (), 983 1.1 mrg (__mmask8) -1, 984 1.1 mrg _MM_FROUND_CUR_DIRECTION); 985 1.1 mrg } 986 1.1 mrg 987 1.1 mrg extern __inline __m256 988 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 989 1.1 mrg _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) 990 1.1 mrg { 991 1.1 mrg return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 992 1.1 mrg (__v8sf) __W, 993 1.1 mrg (__mmask8) __U, 994 1.1 mrg _MM_FROUND_CUR_DIRECTION); 995 1.1 mrg } 996 1.1 mrg 997 1.1 mrg extern __inline __m256 998 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 999 1.1 mrg _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) 1000 1.1 mrg { 1001 1.1 mrg return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1002 1.1 mrg (__v8sf) 1003 1.1 mrg _mm256_setzero_ps (), 1004 1.1 mrg (__mmask8) __U, 1005 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1006 1.1 mrg } 1007 1.1 mrg 1008 1.1 mrg extern __inline __m512d 1009 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1010 1.1 mrg _mm512_cvtepi64_pd (__m512i __A) 1011 1.1 mrg { 1012 1.1 mrg return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1013 1.1 mrg (__v8df) 1014 1.1 mrg _mm512_setzero_pd (), 1015 1.1 mrg (__mmask8) -1, 1016 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1017 1.1 mrg } 1018 1.1 mrg 1019 1.1 mrg extern __inline __m512d 1020 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1021 1.1 mrg _mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) 1022 1.1 mrg { 1023 1.1 mrg return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1024 1.1 mrg (__v8df) __W, 1025 1.1 mrg (__mmask8) __U, 1026 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1027 1.1 mrg } 1028 1.1 mrg 1029 1.1 mrg extern __inline __m512d 1030 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1031 1.1 mrg _mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) 1032 1.1 mrg { 1033 1.1 mrg return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1034 1.1 mrg (__v8df) 1035 1.1 mrg _mm512_setzero_pd (), 1036 1.1 mrg (__mmask8) __U, 1037 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1038 1.1 mrg } 1039 1.1 mrg 1040 1.1 mrg extern __inline __m512d 1041 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1042 1.1 mrg _mm512_cvtepu64_pd (__m512i __A) 1043 1.1 mrg { 1044 1.1 mrg return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1045 1.1 mrg (__v8df) 1046 1.1 mrg _mm512_setzero_pd (), 1047 1.1 mrg (__mmask8) -1, 1048 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1049 1.1 mrg } 1050 1.1 mrg 1051 1.1 mrg extern __inline __m512d 1052 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1053 1.1 mrg _mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) 1054 1.1 mrg { 1055 1.1 mrg return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1056 1.1 mrg (__v8df) __W, 1057 1.1 mrg (__mmask8) __U, 1058 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1059 1.1 mrg } 1060 1.1 mrg 1061 1.1 mrg extern __inline __m512d 1062 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1063 1.1 mrg _mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) 1064 1.1 mrg { 1065 1.1 mrg return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1066 1.1 mrg (__v8df) 1067 1.1 mrg _mm512_setzero_pd (), 1068 1.1 mrg (__mmask8) __U, 1069 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1070 1.1 mrg } 1071 1.1 mrg 1072 1.1 mrg #ifdef __OPTIMIZE__ 1073 1.3 mrg extern __inline __mmask8 1074 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1075 1.3 mrg _kshiftli_mask8 (__mmask8 __A, unsigned int __B) 1076 1.3 mrg { 1077 1.3 mrg return (__mmask8) __builtin_ia32_kshiftliqi ((__mmask8) __A, (__mmask8) __B); 1078 1.3 mrg } 1079 1.3 mrg 1080 1.3 mrg extern __inline __mmask8 1081 1.3 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1082 1.3 mrg _kshiftri_mask8 (__mmask8 __A, unsigned int __B) 1083 1.3 mrg { 1084 1.3 mrg return (__mmask8) __builtin_ia32_kshiftriqi ((__mmask8) __A, (__mmask8) __B); 1085 1.3 mrg } 1086 1.3 mrg 1087 1.1 mrg extern __inline __m512d 1088 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1089 1.1 mrg _mm512_range_pd (__m512d __A, __m512d __B, int __C) 1090 1.1 mrg { 1091 1.1 mrg return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1092 1.1 mrg (__v8df) __B, __C, 1093 1.1 mrg (__v8df) 1094 1.1 mrg _mm512_setzero_pd (), 1095 1.1 mrg (__mmask8) -1, 1096 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1097 1.1 mrg } 1098 1.1 mrg 1099 1.1 mrg extern __inline __m512d 1100 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1101 1.1 mrg _mm512_mask_range_pd (__m512d __W, __mmask8 __U, 1102 1.1 mrg __m512d __A, __m512d __B, int __C) 1103 1.1 mrg { 1104 1.1 mrg return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1105 1.1 mrg (__v8df) __B, __C, 1106 1.1 mrg (__v8df) __W, 1107 1.1 mrg (__mmask8) __U, 1108 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1109 1.1 mrg } 1110 1.1 mrg 1111 1.1 mrg extern __inline __m512d 1112 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1113 1.1 mrg _mm512_maskz_range_pd (__mmask8 __U, __m512d __A, __m512d __B, int __C) 1114 1.1 mrg { 1115 1.1 mrg return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1116 1.1 mrg (__v8df) __B, __C, 1117 1.1 mrg (__v8df) 1118 1.1 mrg _mm512_setzero_pd (), 1119 1.1 mrg (__mmask8) __U, 1120 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1121 1.1 mrg } 1122 1.1 mrg 1123 1.1 mrg extern __inline __m512 1124 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1125 1.1 mrg _mm512_range_ps (__m512 __A, __m512 __B, int __C) 1126 1.1 mrg { 1127 1.1 mrg return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1128 1.1 mrg (__v16sf) __B, __C, 1129 1.1 mrg (__v16sf) 1130 1.1 mrg _mm512_setzero_ps (), 1131 1.1 mrg (__mmask16) -1, 1132 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1133 1.1 mrg } 1134 1.1 mrg 1135 1.1 mrg extern __inline __m512 1136 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1137 1.1 mrg _mm512_mask_range_ps (__m512 __W, __mmask16 __U, 1138 1.1 mrg __m512 __A, __m512 __B, int __C) 1139 1.1 mrg { 1140 1.1 mrg return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1141 1.1 mrg (__v16sf) __B, __C, 1142 1.1 mrg (__v16sf) __W, 1143 1.1 mrg (__mmask16) __U, 1144 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1145 1.1 mrg } 1146 1.1 mrg 1147 1.1 mrg extern __inline __m512 1148 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1149 1.1 mrg _mm512_maskz_range_ps (__mmask16 __U, __m512 __A, __m512 __B, int __C) 1150 1.1 mrg { 1151 1.1 mrg return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1152 1.1 mrg (__v16sf) __B, __C, 1153 1.1 mrg (__v16sf) 1154 1.1 mrg _mm512_setzero_ps (), 1155 1.1 mrg (__mmask16) __U, 1156 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1157 1.1 mrg } 1158 1.1 mrg 1159 1.1 mrg extern __inline __m128d 1160 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1161 1.1 mrg _mm_reduce_sd (__m128d __A, __m128d __B, int __C) 1162 1.1 mrg { 1163 1.4 mrg return (__m128d) __builtin_ia32_reducesd_mask ((__v2df) __A, 1164 1.4 mrg (__v2df) __B, __C, 1165 1.4 mrg (__v2df) _mm_setzero_pd (), 1166 1.4 mrg (__mmask8) -1); 1167 1.4 mrg } 1168 1.4 mrg 1169 1.4 mrg extern __inline __m128d 1170 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1171 1.4 mrg _mm_mask_reduce_sd (__m128d __W, __mmask8 __U, __m128d __A, 1172 1.4 mrg __m128d __B, int __C) 1173 1.4 mrg { 1174 1.4 mrg return (__m128d) __builtin_ia32_reducesd_mask ((__v2df) __A, 1175 1.4 mrg (__v2df) __B, __C, 1176 1.4 mrg (__v2df) __W, 1177 1.4 mrg (__mmask8) __U); 1178 1.4 mrg } 1179 1.4 mrg 1180 1.4 mrg extern __inline __m128d 1181 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1182 1.4 mrg _mm_maskz_reduce_sd (__mmask8 __U, __m128d __A, __m128d __B, int __C) 1183 1.4 mrg { 1184 1.4 mrg return (__m128d) __builtin_ia32_reducesd_mask ((__v2df) __A, 1185 1.4 mrg (__v2df) __B, __C, 1186 1.4 mrg (__v2df) _mm_setzero_pd (), 1187 1.4 mrg (__mmask8) __U); 1188 1.1 mrg } 1189 1.1 mrg 1190 1.1 mrg extern __inline __m128 1191 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1192 1.1 mrg _mm_reduce_ss (__m128 __A, __m128 __B, int __C) 1193 1.1 mrg { 1194 1.4 mrg return (__m128) __builtin_ia32_reducess_mask ((__v4sf) __A, 1195 1.4 mrg (__v4sf) __B, __C, 1196 1.4 mrg (__v4sf) _mm_setzero_ps (), 1197 1.4 mrg (__mmask8) -1); 1198 1.4 mrg } 1199 1.4 mrg 1200 1.4 mrg 1201 1.4 mrg extern __inline __m128 1202 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1203 1.4 mrg _mm_mask_reduce_ss (__m128 __W, __mmask8 __U, __m128 __A, 1204 1.4 mrg __m128 __B, int __C) 1205 1.4 mrg { 1206 1.4 mrg return (__m128) __builtin_ia32_reducess_mask ((__v4sf) __A, 1207 1.4 mrg (__v4sf) __B, __C, 1208 1.4 mrg (__v4sf) __W, 1209 1.4 mrg (__mmask8) __U); 1210 1.4 mrg } 1211 1.4 mrg 1212 1.4 mrg extern __inline __m128 1213 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1214 1.4 mrg _mm_maskz_reduce_ss (__mmask8 __U, __m128 __A, __m128 __B, int __C) 1215 1.4 mrg { 1216 1.4 mrg return (__m128) __builtin_ia32_reducess_mask ((__v4sf) __A, 1217 1.4 mrg (__v4sf) __B, __C, 1218 1.4 mrg (__v4sf) _mm_setzero_ps (), 1219 1.4 mrg (__mmask8) __U); 1220 1.1 mrg } 1221 1.1 mrg 1222 1.1 mrg extern __inline __m128d 1223 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1224 1.1 mrg _mm_range_sd (__m128d __A, __m128d __B, int __C) 1225 1.1 mrg { 1226 1.4 mrg return (__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df) __A, 1227 1.4 mrg (__v2df) __B, __C, 1228 1.4 mrg (__v2df) 1229 1.4 mrg _mm_setzero_pd (), 1230 1.4 mrg (__mmask8) -1, 1231 1.4 mrg _MM_FROUND_CUR_DIRECTION); 1232 1.4 mrg } 1233 1.4 mrg 1234 1.4 mrg extern __inline __m128d 1235 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1236 1.4 mrg _mm_mask_range_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B, int __C) 1237 1.4 mrg { 1238 1.4 mrg return (__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df) __A, 1239 1.1 mrg (__v2df) __B, __C, 1240 1.4 mrg (__v2df) __W, 1241 1.4 mrg (__mmask8) __U, 1242 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1243 1.1 mrg } 1244 1.1 mrg 1245 1.4 mrg extern __inline __m128d 1246 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1247 1.4 mrg _mm_maskz_range_sd (__mmask8 __U, __m128d __A, __m128d __B, int __C) 1248 1.4 mrg { 1249 1.4 mrg return (__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df) __A, 1250 1.4 mrg (__v2df) __B, __C, 1251 1.4 mrg (__v2df) 1252 1.4 mrg _mm_setzero_pd (), 1253 1.4 mrg (__mmask8) __U, 1254 1.4 mrg _MM_FROUND_CUR_DIRECTION); 1255 1.4 mrg } 1256 1.1 mrg 1257 1.1 mrg extern __inline __m128 1258 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1259 1.1 mrg _mm_range_ss (__m128 __A, __m128 __B, int __C) 1260 1.1 mrg { 1261 1.4 mrg return (__m128) __builtin_ia32_rangess128_mask_round ((__v4sf) __A, 1262 1.4 mrg (__v4sf) __B, __C, 1263 1.4 mrg (__v4sf) 1264 1.4 mrg _mm_setzero_ps (), 1265 1.4 mrg (__mmask8) -1, 1266 1.4 mrg _MM_FROUND_CUR_DIRECTION); 1267 1.4 mrg } 1268 1.4 mrg 1269 1.4 mrg extern __inline __m128 1270 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1271 1.4 mrg _mm_mask_range_ss (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B, int __C) 1272 1.4 mrg { 1273 1.4 mrg return (__m128) __builtin_ia32_rangess128_mask_round ((__v4sf) __A, 1274 1.1 mrg (__v4sf) __B, __C, 1275 1.4 mrg (__v4sf) __W, 1276 1.4 mrg (__mmask8) __U, 1277 1.4 mrg _MM_FROUND_CUR_DIRECTION); 1278 1.4 mrg } 1279 1.4 mrg 1280 1.4 mrg 1281 1.4 mrg extern __inline __m128 1282 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1283 1.4 mrg _mm_maskz_range_ss (__mmask8 __U, __m128 __A, __m128 __B, int __C) 1284 1.4 mrg { 1285 1.4 mrg return (__m128) __builtin_ia32_rangess128_mask_round ((__v4sf) __A, 1286 1.4 mrg (__v4sf) __B, __C, 1287 1.4 mrg (__v4sf) 1288 1.4 mrg _mm_setzero_ps (), 1289 1.4 mrg (__mmask8) __U, 1290 1.1 mrg _MM_FROUND_CUR_DIRECTION); 1291 1.1 mrg } 1292 1.1 mrg 1293 1.1 mrg extern __inline __m128d 1294 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1295 1.1 mrg _mm_range_round_sd (__m128d __A, __m128d __B, int __C, const int __R) 1296 1.1 mrg { 1297 1.4 mrg return (__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df) __A, 1298 1.1 mrg (__v2df) __B, __C, 1299 1.4 mrg (__v2df) 1300 1.4 mrg _mm_setzero_pd (), 1301 1.4 mrg (__mmask8) -1, __R); 1302 1.4 mrg } 1303 1.4 mrg 1304 1.4 mrg extern __inline __m128d 1305 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1306 1.4 mrg _mm_mask_range_round_sd (__m128d __W, __mmask8 __U, __m128d __A, __m128d __B, 1307 1.4 mrg int __C, const int __R) 1308 1.4 mrg { 1309 1.4 mrg return (__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df) __A, 1310 1.4 mrg (__v2df) __B, __C, 1311 1.4 mrg (__v2df) __W, 1312 1.4 mrg (__mmask8) __U, __R); 1313 1.4 mrg } 1314 1.4 mrg 1315 1.4 mrg extern __inline __m128d 1316 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1317 1.4 mrg _mm_maskz_range_round_sd (__mmask8 __U, __m128d __A, __m128d __B, int __C, 1318 1.4 mrg const int __R) 1319 1.4 mrg { 1320 1.4 mrg return (__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df) __A, 1321 1.4 mrg (__v2df) __B, __C, 1322 1.4 mrg (__v2df) 1323 1.4 mrg _mm_setzero_pd (), 1324 1.4 mrg (__mmask8) __U, __R); 1325 1.1 mrg } 1326 1.1 mrg 1327 1.1 mrg extern __inline __m128 1328 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1329 1.1 mrg _mm_range_round_ss (__m128 __A, __m128 __B, int __C, const int __R) 1330 1.1 mrg { 1331 1.4 mrg return (__m128) __builtin_ia32_rangess128_mask_round ((__v4sf) __A, 1332 1.4 mrg (__v4sf) __B, __C, 1333 1.4 mrg (__v4sf) 1334 1.4 mrg _mm_setzero_ps (), 1335 1.4 mrg (__mmask8) -1, __R); 1336 1.4 mrg } 1337 1.4 mrg 1338 1.4 mrg extern __inline __m128 1339 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1340 1.4 mrg _mm_mask_range_round_ss (__m128 __W, __mmask8 __U, __m128 __A, __m128 __B, 1341 1.4 mrg int __C, const int __R) 1342 1.4 mrg { 1343 1.4 mrg return (__m128) __builtin_ia32_rangess128_mask_round ((__v4sf) __A, 1344 1.4 mrg (__v4sf) __B, __C, 1345 1.4 mrg (__v4sf) __W, 1346 1.4 mrg (__mmask8) __U, __R); 1347 1.4 mrg } 1348 1.4 mrg 1349 1.4 mrg extern __inline __m128 1350 1.4 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1351 1.4 mrg _mm_maskz_range_round_ss (__mmask8 __U, __m128 __A, __m128 __B, int __C, 1352 1.4 mrg const int __R) 1353 1.4 mrg { 1354 1.4 mrg return (__m128) __builtin_ia32_rangess128_mask_round ((__v4sf) __A, 1355 1.1 mrg (__v4sf) __B, __C, 1356 1.4 mrg (__v4sf) 1357 1.4 mrg _mm_setzero_ps (), 1358 1.4 mrg (__mmask8) __U, __R); 1359 1.1 mrg } 1360 1.1 mrg 1361 1.1 mrg extern __inline __mmask8 1362 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1363 1.1 mrg _mm_fpclass_ss_mask (__m128 __A, const int __imm) 1364 1.1 mrg { 1365 1.1 mrg return (__mmask8) __builtin_ia32_fpclassss ((__v4sf) __A, __imm); 1366 1.1 mrg } 1367 1.1 mrg 1368 1.1 mrg extern __inline __mmask8 1369 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1370 1.1 mrg _mm_fpclass_sd_mask (__m128d __A, const int __imm) 1371 1.1 mrg { 1372 1.1 mrg return (__mmask8) __builtin_ia32_fpclasssd ((__v2df) __A, __imm); 1373 1.1 mrg } 1374 1.1 mrg 1375 1.1 mrg extern __inline __m512i 1376 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1377 1.1 mrg _mm512_cvtt_roundpd_epi64 (__m512d __A, const int __R) 1378 1.1 mrg { 1379 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 1380 1.1 mrg (__v8di) 1381 1.1 mrg _mm512_setzero_si512 (), 1382 1.1 mrg (__mmask8) -1, 1383 1.1 mrg __R); 1384 1.1 mrg } 1385 1.1 mrg 1386 1.1 mrg extern __inline __m512i 1387 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1388 1.1 mrg _mm512_mask_cvtt_roundpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A, 1389 1.1 mrg const int __R) 1390 1.1 mrg { 1391 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 1392 1.1 mrg (__v8di) __W, 1393 1.1 mrg (__mmask8) __U, 1394 1.1 mrg __R); 1395 1.1 mrg } 1396 1.1 mrg 1397 1.1 mrg extern __inline __m512i 1398 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1399 1.1 mrg _mm512_maskz_cvtt_roundpd_epi64 (__mmask8 __U, __m512d __A, 1400 1.1 mrg const int __R) 1401 1.1 mrg { 1402 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 1403 1.1 mrg (__v8di) 1404 1.1 mrg _mm512_setzero_si512 (), 1405 1.1 mrg (__mmask8) __U, 1406 1.1 mrg __R); 1407 1.1 mrg } 1408 1.1 mrg 1409 1.1 mrg extern __inline __m512i 1410 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1411 1.1 mrg _mm512_cvtt_roundpd_epu64 (__m512d __A, const int __R) 1412 1.1 mrg { 1413 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 1414 1.1 mrg (__v8di) 1415 1.1 mrg _mm512_setzero_si512 (), 1416 1.1 mrg (__mmask8) -1, 1417 1.1 mrg __R); 1418 1.1 mrg } 1419 1.1 mrg 1420 1.1 mrg extern __inline __m512i 1421 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1422 1.1 mrg _mm512_mask_cvtt_roundpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A, 1423 1.1 mrg const int __R) 1424 1.1 mrg { 1425 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 1426 1.1 mrg (__v8di) __W, 1427 1.1 mrg (__mmask8) __U, 1428 1.1 mrg __R); 1429 1.1 mrg } 1430 1.1 mrg 1431 1.1 mrg extern __inline __m512i 1432 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1433 1.1 mrg _mm512_maskz_cvtt_roundpd_epu64 (__mmask8 __U, __m512d __A, 1434 1.1 mrg const int __R) 1435 1.1 mrg { 1436 1.1 mrg return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 1437 1.1 mrg (__v8di) 1438 1.1 mrg _mm512_setzero_si512 (), 1439 1.1 mrg (__mmask8) __U, 1440 1.1 mrg __R); 1441 1.1 mrg } 1442 1.1 mrg 1443 1.1 mrg extern __inline __m512i 1444 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1445 1.1 mrg _mm512_cvtt_roundps_epi64 (__m256 __A, const int __R) 1446 1.1 mrg { 1447 1.1 mrg return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 1448 1.1 mrg (__v8di) 1449 1.1 mrg _mm512_setzero_si512 (), 1450 1.1 mrg (__mmask8) -1, 1451 1.1 mrg __R); 1452 1.1 mrg } 1453 1.1 mrg 1454 1.1 mrg extern __inline __m512i 1455 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1456 1.1 mrg _mm512_mask_cvtt_roundps_epi64 (__m512i __W, __mmask8 __U, __m256 __A, 1457 1.1 mrg const int __R) 1458 1.1 mrg { 1459 1.1 mrg return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 1460 1.1 mrg (__v8di) __W, 1461 1.1 mrg (__mmask8) __U, 1462 1.1 mrg __R); 1463 1.1 mrg } 1464 1.1 mrg 1465 1.1 mrg extern __inline __m512i 1466 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1467 1.1 mrg _mm512_maskz_cvtt_roundps_epi64 (__mmask8 __U, __m256 __A, 1468 1.1 mrg const int __R) 1469 1.1 mrg { 1470 1.1 mrg return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 1471 1.1 mrg (__v8di) 1472 1.1 mrg _mm512_setzero_si512 (), 1473 1.1 mrg (__mmask8) __U, 1474 1.1 mrg __R); 1475 1.1 mrg } 1476 1.1 mrg 1477 1.1 mrg extern __inline __m512i 1478 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1479 1.1 mrg _mm512_cvtt_roundps_epu64 (__m256 __A, const int __R) 1480 1.1 mrg { 1481 1.1 mrg return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 1482 1.1 mrg (__v8di) 1483 1.1 mrg _mm512_setzero_si512 (), 1484 1.1 mrg (__mmask8) -1, 1485 1.1 mrg __R); 1486 1.1 mrg } 1487 1.1 mrg 1488 1.1 mrg extern __inline __m512i 1489 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1490 1.1 mrg _mm512_mask_cvtt_roundps_epu64 (__m512i __W, __mmask8 __U, __m256 __A, 1491 1.1 mrg const int __R) 1492 1.1 mrg { 1493 1.1 mrg return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 1494 1.1 mrg (__v8di) __W, 1495 1.1 mrg (__mmask8) __U, 1496 1.1 mrg __R); 1497 1.1 mrg } 1498 1.1 mrg 1499 1.1 mrg extern __inline __m512i 1500 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1501 1.1 mrg _mm512_maskz_cvtt_roundps_epu64 (__mmask8 __U, __m256 __A, 1502 1.1 mrg const int __R) 1503 1.1 mrg { 1504 1.1 mrg return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 1505 1.1 mrg (__v8di) 1506 1.1 mrg _mm512_setzero_si512 (), 1507 1.1 mrg (__mmask8) __U, 1508 1.1 mrg __R); 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_cvt_roundpd_epi64 (__m512d __A, const int __R) 1514 1.1 mrg { 1515 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 1516 1.1 mrg (__v8di) 1517 1.1 mrg _mm512_setzero_si512 (), 1518 1.1 mrg (__mmask8) -1, 1519 1.1 mrg __R); 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_cvt_roundpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A, 1525 1.1 mrg const int __R) 1526 1.1 mrg { 1527 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 1528 1.1 mrg (__v8di) __W, 1529 1.1 mrg (__mmask8) __U, 1530 1.1 mrg __R); 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_cvt_roundpd_epi64 (__mmask8 __U, __m512d __A, 1536 1.1 mrg const int __R) 1537 1.1 mrg { 1538 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 1539 1.1 mrg (__v8di) 1540 1.1 mrg _mm512_setzero_si512 (), 1541 1.1 mrg (__mmask8) __U, 1542 1.1 mrg __R); 1543 1.1 mrg } 1544 1.1 mrg 1545 1.1 mrg extern __inline __m512i 1546 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1547 1.1 mrg _mm512_cvt_roundpd_epu64 (__m512d __A, const int __R) 1548 1.1 mrg { 1549 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 1550 1.1 mrg (__v8di) 1551 1.1 mrg _mm512_setzero_si512 (), 1552 1.1 mrg (__mmask8) -1, 1553 1.1 mrg __R); 1554 1.1 mrg } 1555 1.1 mrg 1556 1.1 mrg extern __inline __m512i 1557 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1558 1.1 mrg _mm512_mask_cvt_roundpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A, 1559 1.1 mrg const int __R) 1560 1.1 mrg { 1561 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 1562 1.1 mrg (__v8di) __W, 1563 1.1 mrg (__mmask8) __U, 1564 1.1 mrg __R); 1565 1.1 mrg } 1566 1.1 mrg 1567 1.1 mrg extern __inline __m512i 1568 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1569 1.1 mrg _mm512_maskz_cvt_roundpd_epu64 (__mmask8 __U, __m512d __A, 1570 1.1 mrg const int __R) 1571 1.1 mrg { 1572 1.1 mrg return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 1573 1.1 mrg (__v8di) 1574 1.1 mrg _mm512_setzero_si512 (), 1575 1.1 mrg (__mmask8) __U, 1576 1.1 mrg __R); 1577 1.1 mrg } 1578 1.1 mrg 1579 1.1 mrg extern __inline __m512i 1580 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1581 1.1 mrg _mm512_cvt_roundps_epi64 (__m256 __A, const int __R) 1582 1.1 mrg { 1583 1.1 mrg return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 1584 1.1 mrg (__v8di) 1585 1.1 mrg _mm512_setzero_si512 (), 1586 1.1 mrg (__mmask8) -1, 1587 1.1 mrg __R); 1588 1.1 mrg } 1589 1.1 mrg 1590 1.1 mrg extern __inline __m512i 1591 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1592 1.1 mrg _mm512_mask_cvt_roundps_epi64 (__m512i __W, __mmask8 __U, __m256 __A, 1593 1.1 mrg const int __R) 1594 1.1 mrg { 1595 1.1 mrg return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 1596 1.1 mrg (__v8di) __W, 1597 1.1 mrg (__mmask8) __U, 1598 1.1 mrg __R); 1599 1.1 mrg } 1600 1.1 mrg 1601 1.1 mrg extern __inline __m512i 1602 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1603 1.1 mrg _mm512_maskz_cvt_roundps_epi64 (__mmask8 __U, __m256 __A, 1604 1.1 mrg const int __R) 1605 1.1 mrg { 1606 1.1 mrg return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 1607 1.1 mrg (__v8di) 1608 1.1 mrg _mm512_setzero_si512 (), 1609 1.1 mrg (__mmask8) __U, 1610 1.1 mrg __R); 1611 1.1 mrg } 1612 1.1 mrg 1613 1.1 mrg extern __inline __m512i 1614 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1615 1.1 mrg _mm512_cvt_roundps_epu64 (__m256 __A, const int __R) 1616 1.1 mrg { 1617 1.1 mrg return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 1618 1.1 mrg (__v8di) 1619 1.1 mrg _mm512_setzero_si512 (), 1620 1.1 mrg (__mmask8) -1, 1621 1.1 mrg __R); 1622 1.1 mrg } 1623 1.1 mrg 1624 1.1 mrg extern __inline __m512i 1625 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1626 1.1 mrg _mm512_mask_cvt_roundps_epu64 (__m512i __W, __mmask8 __U, __m256 __A, 1627 1.1 mrg const int __R) 1628 1.1 mrg { 1629 1.1 mrg return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 1630 1.1 mrg (__v8di) __W, 1631 1.1 mrg (__mmask8) __U, 1632 1.1 mrg __R); 1633 1.1 mrg } 1634 1.1 mrg 1635 1.1 mrg extern __inline __m512i 1636 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1637 1.1 mrg _mm512_maskz_cvt_roundps_epu64 (__mmask8 __U, __m256 __A, 1638 1.1 mrg const int __R) 1639 1.1 mrg { 1640 1.1 mrg return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 1641 1.1 mrg (__v8di) 1642 1.1 mrg _mm512_setzero_si512 (), 1643 1.1 mrg (__mmask8) __U, 1644 1.1 mrg __R); 1645 1.1 mrg } 1646 1.1 mrg 1647 1.1 mrg extern __inline __m256 1648 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1649 1.1 mrg _mm512_cvt_roundepi64_ps (__m512i __A, const int __R) 1650 1.1 mrg { 1651 1.1 mrg return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 1652 1.1 mrg (__v8sf) 1653 1.1 mrg _mm256_setzero_ps (), 1654 1.1 mrg (__mmask8) -1, 1655 1.1 mrg __R); 1656 1.1 mrg } 1657 1.1 mrg 1658 1.1 mrg extern __inline __m256 1659 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1660 1.1 mrg _mm512_mask_cvt_roundepi64_ps (__m256 __W, __mmask8 __U, __m512i __A, 1661 1.1 mrg const int __R) 1662 1.1 mrg { 1663 1.1 mrg return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 1664 1.1 mrg (__v8sf) __W, 1665 1.1 mrg (__mmask8) __U, 1666 1.1 mrg __R); 1667 1.1 mrg } 1668 1.1 mrg 1669 1.1 mrg extern __inline __m256 1670 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1671 1.1 mrg _mm512_maskz_cvt_roundepi64_ps (__mmask8 __U, __m512i __A, 1672 1.1 mrg const int __R) 1673 1.1 mrg { 1674 1.1 mrg return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 1675 1.1 mrg (__v8sf) 1676 1.1 mrg _mm256_setzero_ps (), 1677 1.1 mrg (__mmask8) __U, 1678 1.1 mrg __R); 1679 1.1 mrg } 1680 1.1 mrg 1681 1.1 mrg extern __inline __m256 1682 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1683 1.1 mrg _mm512_cvt_roundepu64_ps (__m512i __A, const int __R) 1684 1.1 mrg { 1685 1.1 mrg return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1686 1.1 mrg (__v8sf) 1687 1.1 mrg _mm256_setzero_ps (), 1688 1.1 mrg (__mmask8) -1, 1689 1.1 mrg __R); 1690 1.1 mrg } 1691 1.1 mrg 1692 1.1 mrg extern __inline __m256 1693 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1694 1.1 mrg _mm512_mask_cvt_roundepu64_ps (__m256 __W, __mmask8 __U, __m512i __A, 1695 1.1 mrg const int __R) 1696 1.1 mrg { 1697 1.1 mrg return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1698 1.1 mrg (__v8sf) __W, 1699 1.1 mrg (__mmask8) __U, 1700 1.1 mrg __R); 1701 1.1 mrg } 1702 1.1 mrg 1703 1.1 mrg extern __inline __m256 1704 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1705 1.1 mrg _mm512_maskz_cvt_roundepu64_ps (__mmask8 __U, __m512i __A, 1706 1.1 mrg const int __R) 1707 1.1 mrg { 1708 1.1 mrg return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1709 1.1 mrg (__v8sf) 1710 1.1 mrg _mm256_setzero_ps (), 1711 1.1 mrg (__mmask8) __U, 1712 1.1 mrg __R); 1713 1.1 mrg } 1714 1.1 mrg 1715 1.1 mrg extern __inline __m512d 1716 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1717 1.1 mrg _mm512_cvt_roundepi64_pd (__m512i __A, const int __R) 1718 1.1 mrg { 1719 1.1 mrg return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1720 1.1 mrg (__v8df) 1721 1.1 mrg _mm512_setzero_pd (), 1722 1.1 mrg (__mmask8) -1, 1723 1.1 mrg __R); 1724 1.1 mrg } 1725 1.1 mrg 1726 1.1 mrg extern __inline __m512d 1727 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1728 1.1 mrg _mm512_mask_cvt_roundepi64_pd (__m512d __W, __mmask8 __U, __m512i __A, 1729 1.1 mrg const int __R) 1730 1.1 mrg { 1731 1.1 mrg return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1732 1.1 mrg (__v8df) __W, 1733 1.1 mrg (__mmask8) __U, 1734 1.1 mrg __R); 1735 1.1 mrg } 1736 1.1 mrg 1737 1.1 mrg extern __inline __m512d 1738 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1739 1.1 mrg _mm512_maskz_cvt_roundepi64_pd (__mmask8 __U, __m512i __A, 1740 1.1 mrg const int __R) 1741 1.1 mrg { 1742 1.1 mrg return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1743 1.1 mrg (__v8df) 1744 1.1 mrg _mm512_setzero_pd (), 1745 1.1 mrg (__mmask8) __U, 1746 1.1 mrg __R); 1747 1.1 mrg } 1748 1.1 mrg 1749 1.1 mrg extern __inline __m512d 1750 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1751 1.1 mrg _mm512_cvt_roundepu64_pd (__m512i __A, const int __R) 1752 1.1 mrg { 1753 1.1 mrg return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1754 1.1 mrg (__v8df) 1755 1.1 mrg _mm512_setzero_pd (), 1756 1.1 mrg (__mmask8) -1, 1757 1.1 mrg __R); 1758 1.1 mrg } 1759 1.1 mrg 1760 1.1 mrg extern __inline __m512d 1761 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1762 1.1 mrg _mm512_mask_cvt_roundepu64_pd (__m512d __W, __mmask8 __U, __m512i __A, 1763 1.1 mrg const int __R) 1764 1.1 mrg { 1765 1.1 mrg return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1766 1.1 mrg (__v8df) __W, 1767 1.1 mrg (__mmask8) __U, 1768 1.1 mrg __R); 1769 1.1 mrg } 1770 1.1 mrg 1771 1.1 mrg extern __inline __m512d 1772 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1773 1.1 mrg _mm512_maskz_cvt_roundepu64_pd (__mmask8 __U, __m512i __A, 1774 1.1 mrg const int __R) 1775 1.1 mrg { 1776 1.1 mrg return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1777 1.1 mrg (__v8df) 1778 1.1 mrg _mm512_setzero_pd (), 1779 1.1 mrg (__mmask8) __U, 1780 1.1 mrg __R); 1781 1.1 mrg } 1782 1.1 mrg 1783 1.1 mrg extern __inline __m512d 1784 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1785 1.1 mrg _mm512_reduce_pd (__m512d __A, int __B) 1786 1.1 mrg { 1787 1.1 mrg return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, 1788 1.1 mrg (__v8df) 1789 1.1 mrg _mm512_setzero_pd (), 1790 1.1 mrg (__mmask8) -1); 1791 1.1 mrg } 1792 1.1 mrg 1793 1.1 mrg extern __inline __m512d 1794 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1795 1.1 mrg _mm512_mask_reduce_pd (__m512d __W, __mmask8 __U, __m512d __A, int __B) 1796 1.1 mrg { 1797 1.1 mrg return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, 1798 1.1 mrg (__v8df) __W, 1799 1.1 mrg (__mmask8) __U); 1800 1.1 mrg } 1801 1.1 mrg 1802 1.1 mrg extern __inline __m512d 1803 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1804 1.1 mrg _mm512_maskz_reduce_pd (__mmask8 __U, __m512d __A, int __B) 1805 1.1 mrg { 1806 1.1 mrg return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, 1807 1.1 mrg (__v8df) 1808 1.1 mrg _mm512_setzero_pd (), 1809 1.1 mrg (__mmask8) __U); 1810 1.1 mrg } 1811 1.1 mrg 1812 1.1 mrg extern __inline __m512 1813 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1814 1.1 mrg _mm512_reduce_ps (__m512 __A, int __B) 1815 1.1 mrg { 1816 1.1 mrg return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, 1817 1.1 mrg (__v16sf) 1818 1.1 mrg _mm512_setzero_ps (), 1819 1.1 mrg (__mmask16) -1); 1820 1.1 mrg } 1821 1.1 mrg 1822 1.1 mrg extern __inline __m512 1823 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1824 1.1 mrg _mm512_mask_reduce_ps (__m512 __W, __mmask16 __U, __m512 __A, int __B) 1825 1.1 mrg { 1826 1.1 mrg return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, 1827 1.1 mrg (__v16sf) __W, 1828 1.1 mrg (__mmask16) __U); 1829 1.1 mrg } 1830 1.1 mrg 1831 1.1 mrg extern __inline __m512 1832 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1833 1.1 mrg _mm512_maskz_reduce_ps (__mmask16 __U, __m512 __A, int __B) 1834 1.1 mrg { 1835 1.1 mrg return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, 1836 1.1 mrg (__v16sf) 1837 1.1 mrg _mm512_setzero_ps (), 1838 1.1 mrg (__mmask16) __U); 1839 1.1 mrg } 1840 1.1 mrg 1841 1.1 mrg extern __inline __m256 1842 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1843 1.1 mrg _mm512_extractf32x8_ps (__m512 __A, const int __imm) 1844 1.1 mrg { 1845 1.1 mrg return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, 1846 1.1 mrg __imm, 1847 1.1 mrg (__v8sf) 1848 1.1 mrg _mm256_setzero_ps (), 1849 1.1 mrg (__mmask8) -1); 1850 1.1 mrg } 1851 1.1 mrg 1852 1.1 mrg extern __inline __m256 1853 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1854 1.1 mrg _mm512_mask_extractf32x8_ps (__m256 __W, __mmask8 __U, __m512 __A, 1855 1.1 mrg const int __imm) 1856 1.1 mrg { 1857 1.1 mrg return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, 1858 1.1 mrg __imm, 1859 1.1 mrg (__v8sf) __W, 1860 1.1 mrg (__mmask8) __U); 1861 1.1 mrg } 1862 1.1 mrg 1863 1.1 mrg extern __inline __m256 1864 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1865 1.1 mrg _mm512_maskz_extractf32x8_ps (__mmask8 __U, __m512 __A, 1866 1.1 mrg const int __imm) 1867 1.1 mrg { 1868 1.1 mrg return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, 1869 1.1 mrg __imm, 1870 1.1 mrg (__v8sf) 1871 1.1 mrg _mm256_setzero_ps (), 1872 1.1 mrg (__mmask8) __U); 1873 1.1 mrg } 1874 1.1 mrg 1875 1.1 mrg extern __inline __m128d 1876 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1877 1.1 mrg _mm512_extractf64x2_pd (__m512d __A, const int __imm) 1878 1.1 mrg { 1879 1.1 mrg return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, 1880 1.1 mrg __imm, 1881 1.1 mrg (__v2df) 1882 1.1 mrg _mm_setzero_pd (), 1883 1.3 mrg (__mmask8) -1); 1884 1.1 mrg } 1885 1.1 mrg 1886 1.1 mrg extern __inline __m128d 1887 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1888 1.1 mrg _mm512_mask_extractf64x2_pd (__m128d __W, __mmask8 __U, __m512d __A, 1889 1.1 mrg const int __imm) 1890 1.1 mrg { 1891 1.1 mrg return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, 1892 1.1 mrg __imm, 1893 1.1 mrg (__v2df) __W, 1894 1.1 mrg (__mmask8) 1895 1.1 mrg __U); 1896 1.1 mrg } 1897 1.1 mrg 1898 1.1 mrg extern __inline __m128d 1899 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1900 1.1 mrg _mm512_maskz_extractf64x2_pd (__mmask8 __U, __m512d __A, 1901 1.1 mrg const int __imm) 1902 1.1 mrg { 1903 1.1 mrg return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, 1904 1.1 mrg __imm, 1905 1.1 mrg (__v2df) 1906 1.1 mrg _mm_setzero_pd (), 1907 1.1 mrg (__mmask8) 1908 1.1 mrg __U); 1909 1.1 mrg } 1910 1.1 mrg 1911 1.1 mrg extern __inline __m256i 1912 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1913 1.1 mrg _mm512_extracti32x8_epi32 (__m512i __A, const int __imm) 1914 1.1 mrg { 1915 1.1 mrg return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, 1916 1.1 mrg __imm, 1917 1.1 mrg (__v8si) 1918 1.1 mrg _mm256_setzero_si256 (), 1919 1.1 mrg (__mmask8) -1); 1920 1.1 mrg } 1921 1.1 mrg 1922 1.1 mrg extern __inline __m256i 1923 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1924 1.1 mrg _mm512_mask_extracti32x8_epi32 (__m256i __W, __mmask8 __U, __m512i __A, 1925 1.1 mrg const int __imm) 1926 1.1 mrg { 1927 1.1 mrg return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, 1928 1.1 mrg __imm, 1929 1.1 mrg (__v8si) __W, 1930 1.1 mrg (__mmask8) __U); 1931 1.1 mrg } 1932 1.1 mrg 1933 1.1 mrg extern __inline __m256i 1934 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1935 1.1 mrg _mm512_maskz_extracti32x8_epi32 (__mmask8 __U, __m512i __A, 1936 1.1 mrg const int __imm) 1937 1.1 mrg { 1938 1.1 mrg return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, 1939 1.1 mrg __imm, 1940 1.1 mrg (__v8si) 1941 1.1 mrg _mm256_setzero_si256 (), 1942 1.1 mrg (__mmask8) __U); 1943 1.1 mrg } 1944 1.1 mrg 1945 1.1 mrg extern __inline __m128i 1946 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1947 1.1 mrg _mm512_extracti64x2_epi64 (__m512i __A, const int __imm) 1948 1.1 mrg { 1949 1.1 mrg return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, 1950 1.1 mrg __imm, 1951 1.1 mrg (__v2di) 1952 1.3 mrg _mm_setzero_si128 (), 1953 1.3 mrg (__mmask8) -1); 1954 1.1 mrg } 1955 1.1 mrg 1956 1.1 mrg extern __inline __m128i 1957 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1958 1.1 mrg _mm512_mask_extracti64x2_epi64 (__m128i __W, __mmask8 __U, __m512i __A, 1959 1.1 mrg const int __imm) 1960 1.1 mrg { 1961 1.1 mrg return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, 1962 1.1 mrg __imm, 1963 1.1 mrg (__v2di) __W, 1964 1.1 mrg (__mmask8) 1965 1.1 mrg __U); 1966 1.1 mrg } 1967 1.1 mrg 1968 1.1 mrg extern __inline __m128i 1969 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1970 1.1 mrg _mm512_maskz_extracti64x2_epi64 (__mmask8 __U, __m512i __A, 1971 1.1 mrg const int __imm) 1972 1.1 mrg { 1973 1.1 mrg return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, 1974 1.1 mrg __imm, 1975 1.1 mrg (__v2di) 1976 1.3 mrg _mm_setzero_si128 (), 1977 1.1 mrg (__mmask8) 1978 1.1 mrg __U); 1979 1.1 mrg } 1980 1.1 mrg 1981 1.1 mrg extern __inline __m512d 1982 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1983 1.1 mrg _mm512_range_round_pd (__m512d __A, __m512d __B, int __C, 1984 1.1 mrg const int __R) 1985 1.1 mrg { 1986 1.1 mrg return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1987 1.1 mrg (__v8df) __B, __C, 1988 1.1 mrg (__v8df) 1989 1.1 mrg _mm512_setzero_pd (), 1990 1.1 mrg (__mmask8) -1, 1991 1.1 mrg __R); 1992 1.1 mrg } 1993 1.1 mrg 1994 1.1 mrg extern __inline __m512d 1995 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1996 1.1 mrg _mm512_mask_range_round_pd (__m512d __W, __mmask8 __U, 1997 1.1 mrg __m512d __A, __m512d __B, int __C, 1998 1.1 mrg const int __R) 1999 1.1 mrg { 2000 1.1 mrg return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 2001 1.1 mrg (__v8df) __B, __C, 2002 1.1 mrg (__v8df) __W, 2003 1.1 mrg (__mmask8) __U, 2004 1.1 mrg __R); 2005 1.1 mrg } 2006 1.1 mrg 2007 1.1 mrg extern __inline __m512d 2008 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2009 1.1 mrg _mm512_maskz_range_round_pd (__mmask8 __U, __m512d __A, __m512d __B, 2010 1.1 mrg int __C, const int __R) 2011 1.1 mrg { 2012 1.1 mrg return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 2013 1.1 mrg (__v8df) __B, __C, 2014 1.1 mrg (__v8df) 2015 1.1 mrg _mm512_setzero_pd (), 2016 1.1 mrg (__mmask8) __U, 2017 1.1 mrg __R); 2018 1.1 mrg } 2019 1.1 mrg 2020 1.1 mrg extern __inline __m512 2021 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2022 1.1 mrg _mm512_range_round_ps (__m512 __A, __m512 __B, int __C, const int __R) 2023 1.1 mrg { 2024 1.1 mrg return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 2025 1.1 mrg (__v16sf) __B, __C, 2026 1.1 mrg (__v16sf) 2027 1.1 mrg _mm512_setzero_ps (), 2028 1.1 mrg (__mmask16) -1, 2029 1.1 mrg __R); 2030 1.1 mrg } 2031 1.1 mrg 2032 1.1 mrg extern __inline __m512 2033 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2034 1.1 mrg _mm512_mask_range_round_ps (__m512 __W, __mmask16 __U, 2035 1.1 mrg __m512 __A, __m512 __B, int __C, 2036 1.1 mrg const int __R) 2037 1.1 mrg { 2038 1.1 mrg return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 2039 1.1 mrg (__v16sf) __B, __C, 2040 1.1 mrg (__v16sf) __W, 2041 1.1 mrg (__mmask16) __U, 2042 1.1 mrg __R); 2043 1.1 mrg } 2044 1.1 mrg 2045 1.1 mrg extern __inline __m512 2046 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2047 1.1 mrg _mm512_maskz_range_round_ps (__mmask16 __U, __m512 __A, __m512 __B, 2048 1.1 mrg int __C, const int __R) 2049 1.1 mrg { 2050 1.1 mrg return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 2051 1.1 mrg (__v16sf) __B, __C, 2052 1.1 mrg (__v16sf) 2053 1.1 mrg _mm512_setzero_ps (), 2054 1.1 mrg (__mmask16) __U, 2055 1.1 mrg __R); 2056 1.1 mrg } 2057 1.1 mrg 2058 1.1 mrg extern __inline __m512i 2059 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2060 1.1 mrg _mm512_inserti32x8 (__m512i __A, __m256i __B, const int __imm) 2061 1.1 mrg { 2062 1.1 mrg return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, 2063 1.1 mrg (__v8si) __B, 2064 1.1 mrg __imm, 2065 1.1 mrg (__v16si) 2066 1.1 mrg _mm512_setzero_si512 (), 2067 1.1 mrg (__mmask16) -1); 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_inserti32x8 (__m512i __W, __mmask16 __U, __m512i __A, 2073 1.1 mrg __m256i __B, const int __imm) 2074 1.1 mrg { 2075 1.1 mrg return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, 2076 1.1 mrg (__v8si) __B, 2077 1.1 mrg __imm, 2078 1.1 mrg (__v16si) __W, 2079 1.1 mrg (__mmask16) __U); 2080 1.1 mrg } 2081 1.1 mrg 2082 1.1 mrg extern __inline __m512i 2083 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2084 1.1 mrg _mm512_maskz_inserti32x8 (__mmask16 __U, __m512i __A, __m256i __B, 2085 1.1 mrg const int __imm) 2086 1.1 mrg { 2087 1.1 mrg return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, 2088 1.1 mrg (__v8si) __B, 2089 1.1 mrg __imm, 2090 1.1 mrg (__v16si) 2091 1.1 mrg _mm512_setzero_si512 (), 2092 1.1 mrg (__mmask16) __U); 2093 1.1 mrg } 2094 1.1 mrg 2095 1.1 mrg extern __inline __m512 2096 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2097 1.1 mrg _mm512_insertf32x8 (__m512 __A, __m256 __B, const int __imm) 2098 1.1 mrg { 2099 1.1 mrg return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, 2100 1.1 mrg (__v8sf) __B, 2101 1.1 mrg __imm, 2102 1.1 mrg (__v16sf) 2103 1.1 mrg _mm512_setzero_ps (), 2104 1.1 mrg (__mmask16) -1); 2105 1.1 mrg } 2106 1.1 mrg 2107 1.1 mrg extern __inline __m512 2108 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2109 1.1 mrg _mm512_mask_insertf32x8 (__m512 __W, __mmask16 __U, __m512 __A, 2110 1.1 mrg __m256 __B, const int __imm) 2111 1.1 mrg { 2112 1.1 mrg return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, 2113 1.1 mrg (__v8sf) __B, 2114 1.1 mrg __imm, 2115 1.1 mrg (__v16sf) __W, 2116 1.1 mrg (__mmask16) __U); 2117 1.1 mrg } 2118 1.1 mrg 2119 1.1 mrg extern __inline __m512 2120 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2121 1.1 mrg _mm512_maskz_insertf32x8 (__mmask16 __U, __m512 __A, __m256 __B, 2122 1.1 mrg const int __imm) 2123 1.1 mrg { 2124 1.1 mrg return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, 2125 1.1 mrg (__v8sf) __B, 2126 1.1 mrg __imm, 2127 1.1 mrg (__v16sf) 2128 1.1 mrg _mm512_setzero_ps (), 2129 1.1 mrg (__mmask16) __U); 2130 1.1 mrg } 2131 1.1 mrg 2132 1.1 mrg extern __inline __m512i 2133 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2134 1.1 mrg _mm512_inserti64x2 (__m512i __A, __m128i __B, const int __imm) 2135 1.1 mrg { 2136 1.1 mrg return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, 2137 1.1 mrg (__v2di) __B, 2138 1.1 mrg __imm, 2139 1.1 mrg (__v8di) 2140 1.1 mrg _mm512_setzero_si512 (), 2141 1.3 mrg (__mmask8) -1); 2142 1.1 mrg } 2143 1.1 mrg 2144 1.1 mrg extern __inline __m512i 2145 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2146 1.1 mrg _mm512_mask_inserti64x2 (__m512i __W, __mmask8 __U, __m512i __A, 2147 1.1 mrg __m128i __B, const int __imm) 2148 1.1 mrg { 2149 1.1 mrg return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, 2150 1.1 mrg (__v2di) __B, 2151 1.1 mrg __imm, 2152 1.1 mrg (__v8di) __W, 2153 1.1 mrg (__mmask8) 2154 1.1 mrg __U); 2155 1.1 mrg } 2156 1.1 mrg 2157 1.1 mrg extern __inline __m512i 2158 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2159 1.1 mrg _mm512_maskz_inserti64x2 (__mmask8 __U, __m512i __A, __m128i __B, 2160 1.1 mrg const int __imm) 2161 1.1 mrg { 2162 1.1 mrg return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, 2163 1.1 mrg (__v2di) __B, 2164 1.1 mrg __imm, 2165 1.1 mrg (__v8di) 2166 1.1 mrg _mm512_setzero_si512 (), 2167 1.1 mrg (__mmask8) 2168 1.1 mrg __U); 2169 1.1 mrg } 2170 1.1 mrg 2171 1.1 mrg extern __inline __m512d 2172 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2173 1.1 mrg _mm512_insertf64x2 (__m512d __A, __m128d __B, const int __imm) 2174 1.1 mrg { 2175 1.1 mrg return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, 2176 1.1 mrg (__v2df) __B, 2177 1.1 mrg __imm, 2178 1.1 mrg (__v8df) 2179 1.1 mrg _mm512_setzero_pd (), 2180 1.3 mrg (__mmask8) -1); 2181 1.1 mrg } 2182 1.1 mrg 2183 1.1 mrg extern __inline __m512d 2184 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2185 1.1 mrg _mm512_mask_insertf64x2 (__m512d __W, __mmask8 __U, __m512d __A, 2186 1.1 mrg __m128d __B, const int __imm) 2187 1.1 mrg { 2188 1.1 mrg return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, 2189 1.1 mrg (__v2df) __B, 2190 1.1 mrg __imm, 2191 1.1 mrg (__v8df) __W, 2192 1.1 mrg (__mmask8) 2193 1.1 mrg __U); 2194 1.1 mrg } 2195 1.1 mrg 2196 1.1 mrg extern __inline __m512d 2197 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2198 1.1 mrg _mm512_maskz_insertf64x2 (__mmask8 __U, __m512d __A, __m128d __B, 2199 1.1 mrg const int __imm) 2200 1.1 mrg { 2201 1.1 mrg return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, 2202 1.1 mrg (__v2df) __B, 2203 1.1 mrg __imm, 2204 1.1 mrg (__v8df) 2205 1.1 mrg _mm512_setzero_pd (), 2206 1.1 mrg (__mmask8) 2207 1.1 mrg __U); 2208 1.1 mrg } 2209 1.1 mrg 2210 1.1 mrg extern __inline __mmask8 2211 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2212 1.1 mrg _mm512_mask_fpclass_pd_mask (__mmask8 __U, __m512d __A, 2213 1.1 mrg const int __imm) 2214 1.1 mrg { 2215 1.1 mrg return (__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) __A, 2216 1.1 mrg __imm, __U); 2217 1.1 mrg } 2218 1.1 mrg 2219 1.1 mrg extern __inline __mmask8 2220 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2221 1.1 mrg _mm512_fpclass_pd_mask (__m512d __A, const int __imm) 2222 1.1 mrg { 2223 1.1 mrg return (__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) __A, 2224 1.1 mrg __imm, 2225 1.1 mrg (__mmask8) -1); 2226 1.1 mrg } 2227 1.1 mrg 2228 1.1 mrg extern __inline __mmask16 2229 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2230 1.1 mrg _mm512_mask_fpclass_ps_mask (__mmask16 __U, __m512 __A, 2231 1.1 mrg const int __imm) 2232 1.1 mrg { 2233 1.1 mrg return (__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) __A, 2234 1.1 mrg __imm, __U); 2235 1.1 mrg } 2236 1.1 mrg 2237 1.1 mrg extern __inline __mmask16 2238 1.1 mrg __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2239 1.1 mrg _mm512_fpclass_ps_mask (__m512 __A, const int __imm) 2240 1.1 mrg { 2241 1.1 mrg return (__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) __A, 2242 1.1 mrg __imm, 2243 1.3 mrg (__mmask16) -1); 2244 1.1 mrg } 2245 1.1 mrg 2246 1.1 mrg #else 2247 1.3 mrg #define _kshiftli_mask8(X, Y) \ 2248 1.3 mrg ((__mmask8) __builtin_ia32_kshiftliqi ((__mmask8)(X), (__mmask8)(Y))) 2249 1.3 mrg 2250 1.3 mrg #define _kshiftri_mask8(X, Y) \ 2251 1.3 mrg ((__mmask8) __builtin_ia32_kshiftriqi ((__mmask8)(X), (__mmask8)(Y))) 2252 1.3 mrg 2253 1.4 mrg #define _mm_range_sd(A, B, C) \ 2254 1.4 mrg ((__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df)(__m128d)(A), \ 2255 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ 2256 1.4 mrg (__mmask8) -1, _MM_FROUND_CUR_DIRECTION)) 2257 1.4 mrg 2258 1.4 mrg #define _mm_mask_range_sd(W, U, A, B, C) \ 2259 1.4 mrg ((__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df)(__m128d)(A), \ 2260 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df)(__m128d)(W), \ 2261 1.4 mrg (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2262 1.4 mrg 2263 1.4 mrg #define _mm_maskz_range_sd(U, A, B, C) \ 2264 1.4 mrg ((__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df)(__m128d)(A), \ 2265 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ 2266 1.4 mrg (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2267 1.1 mrg 2268 1.1 mrg #define _mm_range_ss(A, B, C) \ 2269 1.4 mrg ((__m128) __builtin_ia32_rangess128_mask_round ((__v4sf)(__m128)(A), \ 2270 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ 2271 1.4 mrg (__mmask8) -1, _MM_FROUND_CUR_DIRECTION)) 2272 1.4 mrg 2273 1.4 mrg #define _mm_mask_range_ss(W, U, A, B, C) \ 2274 1.4 mrg ((__m128) __builtin_ia32_rangess128_mask_round ((__v4sf)(__m128)(A), \ 2275 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf)(__m128)(W), \ 2276 1.4 mrg (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2277 1.4 mrg 2278 1.4 mrg #define _mm_maskz_range_ss(U, A, B, C) \ 2279 1.4 mrg ((__m128) __builtin_ia32_rangess128_mask_round ((__v4sf)(__m128)(A), \ 2280 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ 2281 1.4 mrg (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2282 1.4 mrg 2283 1.4 mrg #define _mm_range_round_sd(A, B, C, R) \ 2284 1.4 mrg ((__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df)(__m128d)(A), \ 2285 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ 2286 1.4 mrg (__mmask8) -1, (R))) 2287 1.4 mrg 2288 1.4 mrg #define _mm_mask_range_round_sd(W, U, A, B, C, R) \ 2289 1.4 mrg ((__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df)(__m128d)(A), \ 2290 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df)(__m128d)(W), \ 2291 1.4 mrg (__mmask8)(U), (R))) 2292 1.4 mrg 2293 1.4 mrg #define _mm_maskz_range_round_sd(U, A, B, C, R) \ 2294 1.4 mrg ((__m128d) __builtin_ia32_rangesd128_mask_round ((__v2df)(__m128d)(A), \ 2295 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ 2296 1.4 mrg (__mmask8)(U), (R))) 2297 1.1 mrg 2298 1.1 mrg #define _mm_range_round_ss(A, B, C, R) \ 2299 1.4 mrg ((__m128) __builtin_ia32_rangess128_mask_round ((__v4sf)(__m128)(A), \ 2300 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ 2301 1.4 mrg (__mmask8) -1, (R))) 2302 1.4 mrg 2303 1.4 mrg #define _mm_mask_range_round_ss(W, U, A, B, C, R) \ 2304 1.4 mrg ((__m128) __builtin_ia32_rangess128_mask_round ((__v4sf)(__m128)(A), \ 2305 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf)(__m128)(W), \ 2306 1.4 mrg (__mmask8)(U), (R))) 2307 1.4 mrg 2308 1.4 mrg #define _mm_maskz_range_round_ss(U, A, B, C, R) \ 2309 1.4 mrg ((__m128) __builtin_ia32_rangess128_mask_round ((__v4sf)(__m128)(A), \ 2310 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ 2311 1.4 mrg (__mmask8)(U), (R))) 2312 1.1 mrg 2313 1.1 mrg #define _mm512_cvtt_roundpd_epi64(A, B) \ 2314 1.3 mrg ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di) \ 2315 1.3 mrg _mm512_setzero_si512 (), \ 2316 1.3 mrg -1, (B))) 2317 1.1 mrg 2318 1.1 mrg #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, B) \ 2319 1.3 mrg ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di)(W), (U), (B))) 2320 1.1 mrg 2321 1.1 mrg #define _mm512_maskz_cvtt_roundpd_epi64(U, A, B) \ 2322 1.3 mrg ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2323 1.1 mrg 2324 1.1 mrg #define _mm512_cvtt_roundpd_epu64(A, B) \ 2325 1.3 mrg ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2326 1.1 mrg 2327 1.1 mrg #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, B) \ 2328 1.3 mrg ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2329 1.1 mrg 2330 1.1 mrg #define _mm512_maskz_cvtt_roundpd_epu64(U, A, B) \ 2331 1.3 mrg ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2332 1.1 mrg 2333 1.1 mrg #define _mm512_cvtt_roundps_epi64(A, B) \ 2334 1.3 mrg ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2335 1.1 mrg 2336 1.1 mrg #define _mm512_mask_cvtt_roundps_epi64(W, U, A, B) \ 2337 1.3 mrg ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)(W), (U), (B))) 2338 1.1 mrg 2339 1.1 mrg #define _mm512_maskz_cvtt_roundps_epi64(U, A, B) \ 2340 1.3 mrg ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2341 1.1 mrg 2342 1.1 mrg #define _mm512_cvtt_roundps_epu64(A, B) \ 2343 1.3 mrg ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2344 1.1 mrg 2345 1.1 mrg #define _mm512_mask_cvtt_roundps_epu64(W, U, A, B) \ 2346 1.3 mrg ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2347 1.1 mrg 2348 1.1 mrg #define _mm512_maskz_cvtt_roundps_epu64(U, A, B) \ 2349 1.3 mrg ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2350 1.1 mrg 2351 1.1 mrg #define _mm512_cvt_roundpd_epi64(A, B) \ 2352 1.3 mrg ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2353 1.1 mrg 2354 1.1 mrg #define _mm512_mask_cvt_roundpd_epi64(W, U, A, B) \ 2355 1.3 mrg ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)(W), (U), (B))) 2356 1.1 mrg 2357 1.1 mrg #define _mm512_maskz_cvt_roundpd_epi64(U, A, B) \ 2358 1.3 mrg ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2359 1.1 mrg 2360 1.1 mrg #define _mm512_cvt_roundpd_epu64(A, B) \ 2361 1.3 mrg ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2362 1.1 mrg 2363 1.1 mrg #define _mm512_mask_cvt_roundpd_epu64(W, U, A, B) \ 2364 1.3 mrg ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2365 1.1 mrg 2366 1.1 mrg #define _mm512_maskz_cvt_roundpd_epu64(U, A, B) \ 2367 1.3 mrg ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2368 1.1 mrg 2369 1.1 mrg #define _mm512_cvt_roundps_epi64(A, B) \ 2370 1.3 mrg ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2371 1.1 mrg 2372 1.1 mrg #define _mm512_mask_cvt_roundps_epi64(W, U, A, B) \ 2373 1.3 mrg ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)(W), (U), (B))) 2374 1.1 mrg 2375 1.1 mrg #define _mm512_maskz_cvt_roundps_epi64(U, A, B) \ 2376 1.3 mrg ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2377 1.1 mrg 2378 1.1 mrg #define _mm512_cvt_roundps_epu64(A, B) \ 2379 1.3 mrg ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2380 1.1 mrg 2381 1.1 mrg #define _mm512_mask_cvt_roundps_epu64(W, U, A, B) \ 2382 1.3 mrg ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2383 1.1 mrg 2384 1.1 mrg #define _mm512_maskz_cvt_roundps_epu64(U, A, B) \ 2385 1.3 mrg ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2386 1.1 mrg 2387 1.1 mrg #define _mm512_cvt_roundepi64_ps(A, B) \ 2388 1.3 mrg ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), -1, (B))) 2389 1.1 mrg 2390 1.1 mrg #define _mm512_mask_cvt_roundepi64_ps(W, U, A, B) \ 2391 1.3 mrg ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (W), (U), (B))) 2392 1.1 mrg 2393 1.1 mrg #define _mm512_maskz_cvt_roundepi64_ps(U, A, B) \ 2394 1.3 mrg ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), (U), (B))) 2395 1.1 mrg 2396 1.1 mrg #define _mm512_cvt_roundepu64_ps(A, B) \ 2397 1.3 mrg ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), -1, (B))) 2398 1.1 mrg 2399 1.1 mrg #define _mm512_mask_cvt_roundepu64_ps(W, U, A, B) \ 2400 1.3 mrg ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (W), (U), (B))) 2401 1.1 mrg 2402 1.1 mrg #define _mm512_maskz_cvt_roundepu64_ps(U, A, B) \ 2403 1.3 mrg ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), (U), (B))) 2404 1.1 mrg 2405 1.1 mrg #define _mm512_cvt_roundepi64_pd(A, B) \ 2406 1.3 mrg ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), -1, (B))) 2407 1.1 mrg 2408 1.1 mrg #define _mm512_mask_cvt_roundepi64_pd(W, U, A, B) \ 2409 1.3 mrg ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (W), (U), (B))) 2410 1.1 mrg 2411 1.1 mrg #define _mm512_maskz_cvt_roundepi64_pd(U, A, B) \ 2412 1.3 mrg ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), (U), (B))) 2413 1.1 mrg 2414 1.1 mrg #define _mm512_cvt_roundepu64_pd(A, B) \ 2415 1.3 mrg ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), -1, (B))) 2416 1.1 mrg 2417 1.1 mrg #define _mm512_mask_cvt_roundepu64_pd(W, U, A, B) \ 2418 1.3 mrg ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (W), (U), (B))) 2419 1.1 mrg 2420 1.1 mrg #define _mm512_maskz_cvt_roundepu64_pd(U, A, B) \ 2421 1.3 mrg ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), (U), (B))) 2422 1.1 mrg 2423 1.1 mrg #define _mm512_reduce_pd(A, B) \ 2424 1.1 mrg ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ 2425 1.3 mrg (int)(B), (__v8df)_mm512_setzero_pd (), (__mmask8)-1)) 2426 1.1 mrg 2427 1.1 mrg #define _mm512_mask_reduce_pd(W, U, A, B) \ 2428 1.1 mrg ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ 2429 1.1 mrg (int)(B), (__v8df)(__m512d)(W), (__mmask8)(U))) 2430 1.1 mrg 2431 1.1 mrg #define _mm512_maskz_reduce_pd(U, A, B) \ 2432 1.1 mrg ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ 2433 1.3 mrg (int)(B), (__v8df)_mm512_setzero_pd (), (__mmask8)(U))) 2434 1.1 mrg 2435 1.1 mrg #define _mm512_reduce_ps(A, B) \ 2436 1.1 mrg ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ 2437 1.3 mrg (int)(B), (__v16sf)_mm512_setzero_ps (), (__mmask16)-1)) 2438 1.1 mrg 2439 1.1 mrg #define _mm512_mask_reduce_ps(W, U, A, B) \ 2440 1.1 mrg ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ 2441 1.1 mrg (int)(B), (__v16sf)(__m512)(W), (__mmask16)(U))) 2442 1.1 mrg 2443 1.1 mrg #define _mm512_maskz_reduce_ps(U, A, B) \ 2444 1.1 mrg ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ 2445 1.3 mrg (int)(B), (__v16sf)_mm512_setzero_ps (), (__mmask16)(U))) 2446 1.1 mrg 2447 1.1 mrg #define _mm512_extractf32x8_ps(X, C) \ 2448 1.1 mrg ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ 2449 1.3 mrg (int) (C), (__v8sf)(__m256) _mm256_setzero_ps (), (__mmask8)-1)) 2450 1.1 mrg 2451 1.1 mrg #define _mm512_mask_extractf32x8_ps(W, U, X, C) \ 2452 1.1 mrg ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ 2453 1.1 mrg (int) (C), (__v8sf)(__m256) (W), (__mmask8) (U))) 2454 1.1 mrg 2455 1.1 mrg #define _mm512_maskz_extractf32x8_ps(U, X, C) \ 2456 1.1 mrg ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ 2457 1.3 mrg (int) (C), (__v8sf)(__m256) _mm256_setzero_ps (), (__mmask8) (U))) 2458 1.1 mrg 2459 1.1 mrg #define _mm512_extractf64x2_pd(X, C) \ 2460 1.1 mrg ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ 2461 1.3 mrg (int) (C), (__v2df)(__m128d) _mm_setzero_pd (), (__mmask8)-1)) 2462 1.1 mrg 2463 1.1 mrg #define _mm512_mask_extractf64x2_pd(W, U, X, C) \ 2464 1.1 mrg ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ 2465 1.1 mrg (int) (C), (__v2df)(__m128d) (W), (__mmask8) (U))) 2466 1.1 mrg 2467 1.1 mrg #define _mm512_maskz_extractf64x2_pd(U, X, C) \ 2468 1.1 mrg ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ 2469 1.3 mrg (int) (C), (__v2df)(__m128d) _mm_setzero_pd (), (__mmask8) (U))) 2470 1.1 mrg 2471 1.1 mrg #define _mm512_extracti32x8_epi32(X, C) \ 2472 1.1 mrg ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ 2473 1.3 mrg (int) (C), (__v8si)(__m256i) _mm256_setzero_si256 (), (__mmask8)-1)) 2474 1.1 mrg 2475 1.1 mrg #define _mm512_mask_extracti32x8_epi32(W, U, X, C) \ 2476 1.1 mrg ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ 2477 1.1 mrg (int) (C), (__v8si)(__m256i) (W), (__mmask8) (U))) 2478 1.1 mrg 2479 1.1 mrg #define _mm512_maskz_extracti32x8_epi32(U, X, C) \ 2480 1.1 mrg ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ 2481 1.3 mrg (int) (C), (__v8si)(__m256i) _mm256_setzero_si256 (), (__mmask8) (U))) 2482 1.1 mrg 2483 1.1 mrg #define _mm512_extracti64x2_epi64(X, C) \ 2484 1.1 mrg ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ 2485 1.3 mrg (int) (C), (__v2di)(__m128i) _mm_setzero_si128 (), (__mmask8)-1)) 2486 1.1 mrg 2487 1.1 mrg #define _mm512_mask_extracti64x2_epi64(W, U, X, C) \ 2488 1.1 mrg ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ 2489 1.1 mrg (int) (C), (__v2di)(__m128i) (W), (__mmask8) (U))) 2490 1.1 mrg 2491 1.1 mrg #define _mm512_maskz_extracti64x2_epi64(U, X, C) \ 2492 1.1 mrg ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ 2493 1.3 mrg (int) (C), (__v2di)(__m128i) _mm_setzero_si128 (), (__mmask8) (U))) 2494 1.1 mrg 2495 1.1 mrg #define _mm512_range_pd(A, B, C) \ 2496 1.1 mrg ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2497 1.1 mrg (__v8df)(__m512d)(B), (int)(C), \ 2498 1.3 mrg (__v8df)_mm512_setzero_pd (), (__mmask8)-1, _MM_FROUND_CUR_DIRECTION)) 2499 1.1 mrg 2500 1.1 mrg #define _mm512_mask_range_pd(W, U, A, B, C) \ 2501 1.1 mrg ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2502 1.1 mrg (__v8df)(__m512d)(B), (int)(C), \ 2503 1.1 mrg (__v8df)(__m512d)(W), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2504 1.1 mrg 2505 1.1 mrg #define _mm512_maskz_range_pd(U, A, B, C) \ 2506 1.1 mrg ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2507 1.1 mrg (__v8df)(__m512d)(B), (int)(C), \ 2508 1.3 mrg (__v8df)_mm512_setzero_pd (), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2509 1.1 mrg 2510 1.1 mrg #define _mm512_range_ps(A, B, C) \ 2511 1.1 mrg ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2512 1.1 mrg (__v16sf)(__m512)(B), (int)(C), \ 2513 1.3 mrg (__v16sf)_mm512_setzero_ps (), (__mmask16)-1, _MM_FROUND_CUR_DIRECTION)) 2514 1.1 mrg 2515 1.1 mrg #define _mm512_mask_range_ps(W, U, A, B, C) \ 2516 1.1 mrg ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2517 1.1 mrg (__v16sf)(__m512)(B), (int)(C), \ 2518 1.1 mrg (__v16sf)(__m512)(W), (__mmask16)(U), _MM_FROUND_CUR_DIRECTION)) 2519 1.1 mrg 2520 1.1 mrg #define _mm512_maskz_range_ps(U, A, B, C) \ 2521 1.1 mrg ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2522 1.1 mrg (__v16sf)(__m512)(B), (int)(C), \ 2523 1.3 mrg (__v16sf)_mm512_setzero_ps (), (__mmask16)(U), _MM_FROUND_CUR_DIRECTION)) 2524 1.1 mrg 2525 1.1 mrg #define _mm512_range_round_pd(A, B, C, R) \ 2526 1.1 mrg ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2527 1.1 mrg (__v8df)(__m512d)(B), (int)(C), \ 2528 1.3 mrg (__v8df)_mm512_setzero_pd (), (__mmask8)-1, (R))) 2529 1.1 mrg 2530 1.1 mrg #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ 2531 1.1 mrg ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2532 1.1 mrg (__v8df)(__m512d)(B), (int)(C), \ 2533 1.1 mrg (__v8df)(__m512d)(W), (__mmask8)(U), (R))) 2534 1.1 mrg 2535 1.1 mrg #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ 2536 1.1 mrg ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2537 1.1 mrg (__v8df)(__m512d)(B), (int)(C), \ 2538 1.3 mrg (__v8df)_mm512_setzero_pd (), (__mmask8)(U), (R))) 2539 1.1 mrg 2540 1.1 mrg #define _mm512_range_round_ps(A, B, C, R) \ 2541 1.1 mrg ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2542 1.1 mrg (__v16sf)(__m512)(B), (int)(C), \ 2543 1.3 mrg (__v16sf)_mm512_setzero_ps (), (__mmask16)-1, (R))) 2544 1.1 mrg 2545 1.1 mrg #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ 2546 1.1 mrg ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2547 1.1 mrg (__v16sf)(__m512)(B), (int)(C), \ 2548 1.1 mrg (__v16sf)(__m512)(W), (__mmask16)(U), (R))) 2549 1.1 mrg 2550 1.1 mrg #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ 2551 1.1 mrg ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2552 1.1 mrg (__v16sf)(__m512)(B), (int)(C), \ 2553 1.3 mrg (__v16sf)_mm512_setzero_ps (), (__mmask16)(U), (R))) 2554 1.1 mrg 2555 1.1 mrg #define _mm512_insertf64x2(X, Y, C) \ 2556 1.1 mrg ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ 2557 1.1 mrg (__v2df)(__m128d) (Y), (int) (C), (__v8df)(__m512d) (X), \ 2558 1.1 mrg (__mmask8)-1)) 2559 1.1 mrg 2560 1.1 mrg #define _mm512_mask_insertf64x2(W, U, X, Y, C) \ 2561 1.1 mrg ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ 2562 1.1 mrg (__v2df)(__m128d) (Y), (int) (C), (__v8df)(__m512d) (W), \ 2563 1.1 mrg (__mmask8) (U))) 2564 1.1 mrg 2565 1.1 mrg #define _mm512_maskz_insertf64x2(U, X, Y, C) \ 2566 1.1 mrg ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ 2567 1.1 mrg (__v2df)(__m128d) (Y), (int) (C), \ 2568 1.3 mrg (__v8df)(__m512d) _mm512_setzero_pd (), (__mmask8) (U))) 2569 1.1 mrg 2570 1.1 mrg #define _mm512_inserti64x2(X, Y, C) \ 2571 1.1 mrg ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ 2572 1.1 mrg (__v2di)(__m128i) (Y), (int) (C), (__v8di)(__m512i) (X), (__mmask8)-1)) 2573 1.1 mrg 2574 1.1 mrg #define _mm512_mask_inserti64x2(W, U, X, Y, C) \ 2575 1.1 mrg ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ 2576 1.1 mrg (__v2di)(__m128i) (Y), (int) (C), (__v8di)(__m512i) (W), \ 2577 1.1 mrg (__mmask8) (U))) 2578 1.1 mrg 2579 1.1 mrg #define _mm512_maskz_inserti64x2(U, X, Y, C) \ 2580 1.1 mrg ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ 2581 1.1 mrg (__v2di)(__m128i) (Y), (int) (C), \ 2582 1.1 mrg (__v8di)(__m512i) _mm512_setzero_si512 (), (__mmask8) (U))) 2583 1.1 mrg 2584 1.1 mrg #define _mm512_insertf32x8(X, Y, C) \ 2585 1.1 mrg ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ 2586 1.1 mrg (__v8sf)(__m256) (Y), (int) (C),\ 2587 1.3 mrg (__v16sf)(__m512)_mm512_setzero_ps (),\ 2588 1.1 mrg (__mmask16)-1)) 2589 1.1 mrg 2590 1.1 mrg #define _mm512_mask_insertf32x8(W, U, X, Y, C) \ 2591 1.1 mrg ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ 2592 1.1 mrg (__v8sf)(__m256) (Y), (int) (C),\ 2593 1.1 mrg (__v16sf)(__m512)(W),\ 2594 1.1 mrg (__mmask16)(U))) 2595 1.1 mrg 2596 1.1 mrg #define _mm512_maskz_insertf32x8(U, X, Y, C) \ 2597 1.1 mrg ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ 2598 1.1 mrg (__v8sf)(__m256) (Y), (int) (C),\ 2599 1.3 mrg (__v16sf)(__m512)_mm512_setzero_ps (),\ 2600 1.1 mrg (__mmask16)(U))) 2601 1.1 mrg 2602 1.1 mrg #define _mm512_inserti32x8(X, Y, C) \ 2603 1.1 mrg ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ 2604 1.1 mrg (__v8si)(__m256i) (Y), (int) (C),\ 2605 1.1 mrg (__v16si)(__m512i)_mm512_setzero_si512 (),\ 2606 1.1 mrg (__mmask16)-1)) 2607 1.1 mrg 2608 1.1 mrg #define _mm512_mask_inserti32x8(W, U, X, Y, C) \ 2609 1.1 mrg ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ 2610 1.1 mrg (__v8si)(__m256i) (Y), (int) (C),\ 2611 1.1 mrg (__v16si)(__m512i)(W),\ 2612 1.1 mrg (__mmask16)(U))) 2613 1.1 mrg 2614 1.1 mrg #define _mm512_maskz_inserti32x8(U, X, Y, C) \ 2615 1.1 mrg ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ 2616 1.1 mrg (__v8si)(__m256i) (Y), (int) (C),\ 2617 1.1 mrg (__v16si)(__m512i)_mm512_setzero_si512 (),\ 2618 1.1 mrg (__mmask16)(U))) 2619 1.1 mrg 2620 1.1 mrg #define _mm_fpclass_ss_mask(X, C) \ 2621 1.1 mrg ((__mmask8) __builtin_ia32_fpclassss ((__v4sf) (__m128) (X), (int) (C))) \ 2622 1.1 mrg 2623 1.1 mrg #define _mm_fpclass_sd_mask(X, C) \ 2624 1.1 mrg ((__mmask8) __builtin_ia32_fpclasssd ((__v2df) (__m128d) (X), (int) (C))) \ 2625 1.1 mrg 2626 1.1 mrg #define _mm512_mask_fpclass_pd_mask(u, X, C) \ 2627 1.1 mrg ((__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) (__m512d) (X), \ 2628 1.1 mrg (int) (C), (__mmask8)(u))) 2629 1.1 mrg 2630 1.1 mrg #define _mm512_mask_fpclass_ps_mask(u, x, c) \ 2631 1.1 mrg ((__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) (__m512) (x),\ 2632 1.1 mrg (int) (c),(__mmask8)(u))) 2633 1.1 mrg 2634 1.1 mrg #define _mm512_fpclass_pd_mask(X, C) \ 2635 1.1 mrg ((__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) (__m512d) (X), \ 2636 1.1 mrg (int) (C), (__mmask8)-1)) 2637 1.1 mrg 2638 1.1 mrg #define _mm512_fpclass_ps_mask(x, c) \ 2639 1.1 mrg ((__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) (__m512) (x),\ 2640 1.1 mrg (int) (c),(__mmask8)-1)) 2641 1.1 mrg 2642 1.1 mrg #define _mm_reduce_sd(A, B, C) \ 2643 1.4 mrg ((__m128d) __builtin_ia32_reducesd_mask ((__v2df)(__m128d)(A), \ 2644 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ 2645 1.4 mrg (__mmask8)-1)) 2646 1.4 mrg 2647 1.4 mrg #define _mm_mask_reduce_sd(W, U, A, B, C) \ 2648 1.4 mrg ((__m128d) __builtin_ia32_reducesd_mask ((__v2df)(__m128d)(A), \ 2649 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df)(__m128d)(W), (__mmask8)(U))) 2650 1.4 mrg 2651 1.4 mrg #define _mm_maskz_reduce_sd(U, A, B, C) \ 2652 1.4 mrg ((__m128d) __builtin_ia32_reducesd_mask ((__v2df)(__m128d)(A), \ 2653 1.4 mrg (__v2df)(__m128d)(B), (int)(C), (__v2df) _mm_setzero_pd (), \ 2654 1.4 mrg (__mmask8)(U))) 2655 1.1 mrg 2656 1.1 mrg #define _mm_reduce_ss(A, B, C) \ 2657 1.4 mrg ((__m128) __builtin_ia32_reducess_mask ((__v4sf)(__m128)(A), \ 2658 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ 2659 1.4 mrg (__mmask8)-1)) 2660 1.4 mrg 2661 1.4 mrg #define _mm_mask_reduce_ss(W, U, A, B, C) \ 2662 1.4 mrg ((__m128) __builtin_ia32_reducess_mask ((__v4sf)(__m128)(A), \ 2663 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf)(__m128)(W), (__mmask8)(U))) 2664 1.4 mrg 2665 1.4 mrg #define _mm_maskz_reduce_ss(U, A, B, C) \ 2666 1.4 mrg ((__m128) __builtin_ia32_reducess_mask ((__v4sf)(__m128)(A), \ 2667 1.4 mrg (__v4sf)(__m128)(B), (int)(C), (__v4sf) _mm_setzero_ps (), \ 2668 1.4 mrg (__mmask8)(U))) 2669 1.4 mrg 2670 1.4 mrg 2671 1.1 mrg 2672 1.1 mrg #endif 2673 1.1 mrg 2674 1.1 mrg #ifdef __DISABLE_AVX512DQ__ 2675 1.1 mrg #undef __DISABLE_AVX512DQ__ 2676 1.1 mrg #pragma GCC pop_options 2677 1.1 mrg #endif /* __DISABLE_AVX512DQ__ */ 2678 1.1 mrg 2679 1.1 mrg #endif /* _AVX512DQINTRIN_H_INCLUDED */ 2680
Indexes created Sat Apr 25 00:22:52 UTC 2026