avx512dqintrin.h revision 1.3
1 /* Copyright (C) 2014-2017 Free Software Foundation, Inc. 2 3 This file is part of GCC. 4 5 GCC is free software; you can redistribute it and/or modify 6 it under the terms of the GNU General Public License as published by 7 the Free Software Foundation; either version 3, or (at your option) 8 any later version. 9 10 GCC is distributed in the hope that it will be useful, 11 but WITHOUT ANY WARRANTY; without even the implied warranty of 12 MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the 13 GNU General Public License for more details. 14 15 Under Section 7 of GPL version 3, you are granted additional 16 permissions described in the GCC Runtime Library Exception, version 17 3.1, as published by the Free Software Foundation. 18 19 You should have received a copy of the GNU General Public License and 20 a copy of the GCC Runtime Library Exception along with this program; 21 see the files COPYING3 and COPYING.RUNTIME respectively. If not, see 22 <http://www.gnu.org/licenses/>. */ 23 24 #ifndef _IMMINTRIN_H_INCLUDED 25 #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead." 26 #endif 27 28 #ifndef _AVX512DQINTRIN_H_INCLUDED 29 #define _AVX512DQINTRIN_H_INCLUDED 30 31 #ifndef __AVX512DQ__ 32 #pragma GCC push_options 33 #pragma GCC target("avx512dq") 34 #define __DISABLE_AVX512DQ__ 35 #endif /* __AVX512DQ__ */ 36 37 extern __inline unsigned char 38 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 39 _ktest_mask8_u8 (__mmask8 __A, __mmask8 __B, unsigned char *__CF) 40 { 41 *__CF = (unsigned char) __builtin_ia32_ktestcqi (__A, __B); 42 return (unsigned char) __builtin_ia32_ktestzqi (__A, __B); 43 } 44 45 extern __inline unsigned char 46 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 47 _ktestz_mask8_u8 (__mmask8 __A, __mmask8 __B) 48 { 49 return (unsigned char) __builtin_ia32_ktestzqi (__A, __B); 50 } 51 52 extern __inline unsigned char 53 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 54 _ktestc_mask8_u8 (__mmask8 __A, __mmask8 __B) 55 { 56 return (unsigned char) __builtin_ia32_ktestcqi (__A, __B); 57 } 58 59 extern __inline unsigned char 60 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 61 _ktest_mask16_u8 (__mmask16 __A, __mmask16 __B, unsigned char *__CF) 62 { 63 *__CF = (unsigned char) __builtin_ia32_ktestchi (__A, __B); 64 return (unsigned char) __builtin_ia32_ktestzhi (__A, __B); 65 } 66 67 extern __inline unsigned char 68 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 69 _ktestz_mask16_u8 (__mmask16 __A, __mmask16 __B) 70 { 71 return (unsigned char) __builtin_ia32_ktestzhi (__A, __B); 72 } 73 74 extern __inline unsigned char 75 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 76 _ktestc_mask16_u8 (__mmask16 __A, __mmask16 __B) 77 { 78 return (unsigned char) __builtin_ia32_ktestchi (__A, __B); 79 } 80 81 extern __inline unsigned char 82 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 83 _kortest_mask8_u8 (__mmask8 __A, __mmask8 __B, unsigned char *__CF) 84 { 85 *__CF = (unsigned char) __builtin_ia32_kortestcqi (__A, __B); 86 return (unsigned char) __builtin_ia32_kortestzqi (__A, __B); 87 } 88 89 extern __inline unsigned char 90 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 91 _kortestz_mask8_u8 (__mmask8 __A, __mmask8 __B) 92 { 93 return (unsigned char) __builtin_ia32_kortestzqi (__A, __B); 94 } 95 96 extern __inline unsigned char 97 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 98 _kortestc_mask8_u8 (__mmask8 __A, __mmask8 __B) 99 { 100 return (unsigned char) __builtin_ia32_kortestcqi (__A, __B); 101 } 102 103 extern __inline __mmask8 104 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 105 _kadd_mask8 (__mmask8 __A, __mmask8 __B) 106 { 107 return (__mmask8) __builtin_ia32_kaddqi ((__mmask8) __A, (__mmask8) __B); 108 } 109 110 extern __inline __mmask16 111 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 112 _kadd_mask16 (__mmask16 __A, __mmask16 __B) 113 { 114 return (__mmask16) __builtin_ia32_kaddhi ((__mmask16) __A, (__mmask16) __B); 115 } 116 117 extern __inline unsigned int 118 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 119 _cvtmask8_u32 (__mmask8 __A) 120 { 121 return (unsigned int) __builtin_ia32_kmovb ((__mmask8 ) __A); 122 } 123 124 extern __inline __mmask8 125 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 126 _cvtu32_mask8 (unsigned int __A) 127 { 128 return (__mmask8) __builtin_ia32_kmovb ((__mmask8) __A); 129 } 130 131 extern __inline __mmask8 132 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 133 _load_mask8 (__mmask8 *__A) 134 { 135 return (__mmask8) __builtin_ia32_kmovb (*(__mmask8 *) __A); 136 } 137 138 extern __inline void 139 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 140 _store_mask8 (__mmask8 *__A, __mmask8 __B) 141 { 142 *(__mmask8 *) __A = __builtin_ia32_kmovb (__B); 143 } 144 145 extern __inline __mmask8 146 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 147 _knot_mask8 (__mmask8 __A) 148 { 149 return (__mmask8) __builtin_ia32_knotqi ((__mmask8) __A); 150 } 151 152 extern __inline __mmask8 153 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 154 _kor_mask8 (__mmask8 __A, __mmask8 __B) 155 { 156 return (__mmask8) __builtin_ia32_korqi ((__mmask8) __A, (__mmask8) __B); 157 } 158 159 extern __inline __mmask8 160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 161 _kxnor_mask8 (__mmask8 __A, __mmask8 __B) 162 { 163 return (__mmask8) __builtin_ia32_kxnorqi ((__mmask8) __A, (__mmask8) __B); 164 } 165 166 extern __inline __mmask8 167 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 168 _kxor_mask8 (__mmask8 __A, __mmask8 __B) 169 { 170 return (__mmask8) __builtin_ia32_kxorqi ((__mmask8) __A, (__mmask8) __B); 171 } 172 173 extern __inline __mmask8 174 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 175 _kand_mask8 (__mmask8 __A, __mmask8 __B) 176 { 177 return (__mmask8) __builtin_ia32_kandqi ((__mmask8) __A, (__mmask8) __B); 178 } 179 180 extern __inline __mmask8 181 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 182 _kandn_mask8 (__mmask8 __A, __mmask8 __B) 183 { 184 return (__mmask8) __builtin_ia32_kandnqi ((__mmask8) __A, (__mmask8) __B); 185 } 186 187 extern __inline __m512d 188 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 189 _mm512_broadcast_f64x2 (__m128d __A) 190 { 191 return (__m512d) 192 __builtin_ia32_broadcastf64x2_512_mask ((__v2df) __A, 193 _mm512_undefined_pd (), 194 (__mmask8) -1); 195 } 196 197 extern __inline __m512d 198 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 199 _mm512_mask_broadcast_f64x2 (__m512d __O, __mmask8 __M, __m128d __A) 200 { 201 return (__m512d) __builtin_ia32_broadcastf64x2_512_mask ((__v2df) 202 __A, 203 (__v8df) 204 __O, __M); 205 } 206 207 extern __inline __m512d 208 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 209 _mm512_maskz_broadcast_f64x2 (__mmask8 __M, __m128d __A) 210 { 211 return (__m512d) __builtin_ia32_broadcastf64x2_512_mask ((__v2df) 212 __A, 213 (__v8df) 214 _mm512_setzero_ps (), 215 __M); 216 } 217 218 extern __inline __m512i 219 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 220 _mm512_broadcast_i64x2 (__m128i __A) 221 { 222 return (__m512i) 223 __builtin_ia32_broadcasti64x2_512_mask ((__v2di) __A, 224 _mm512_undefined_epi32 (), 225 (__mmask8) -1); 226 } 227 228 extern __inline __m512i 229 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 230 _mm512_mask_broadcast_i64x2 (__m512i __O, __mmask8 __M, __m128i __A) 231 { 232 return (__m512i) __builtin_ia32_broadcasti64x2_512_mask ((__v2di) 233 __A, 234 (__v8di) 235 __O, __M); 236 } 237 238 extern __inline __m512i 239 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 240 _mm512_maskz_broadcast_i64x2 (__mmask8 __M, __m128i __A) 241 { 242 return (__m512i) __builtin_ia32_broadcasti64x2_512_mask ((__v2di) 243 __A, 244 (__v8di) 245 _mm512_setzero_si512 (), 246 __M); 247 } 248 249 extern __inline __m512 250 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 251 _mm512_broadcast_f32x2 (__m128 __A) 252 { 253 return (__m512) 254 __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, 255 (__v16sf)_mm512_undefined_ps (), 256 (__mmask16) -1); 257 } 258 259 extern __inline __m512 260 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 261 _mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A) 262 { 263 return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, 264 (__v16sf) 265 __O, __M); 266 } 267 268 extern __inline __m512 269 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 270 _mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A) 271 { 272 return (__m512) __builtin_ia32_broadcastf32x2_512_mask ((__v4sf) __A, 273 (__v16sf) 274 _mm512_setzero_ps (), 275 __M); 276 } 277 278 extern __inline __m512i 279 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 280 _mm512_broadcast_i32x2 (__m128i __A) 281 { 282 return (__m512i) 283 __builtin_ia32_broadcasti32x2_512_mask ((__v4si) __A, 284 (__v16si) 285 _mm512_undefined_epi32 (), 286 (__mmask16) -1); 287 } 288 289 extern __inline __m512i 290 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 291 _mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A) 292 { 293 return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) 294 __A, 295 (__v16si) 296 __O, __M); 297 } 298 299 extern __inline __m512i 300 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 301 _mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A) 302 { 303 return (__m512i) __builtin_ia32_broadcasti32x2_512_mask ((__v4si) 304 __A, 305 (__v16si) 306 _mm512_setzero_si512 (), 307 __M); 308 } 309 310 extern __inline __m512 311 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 312 _mm512_broadcast_f32x8 (__m256 __A) 313 { 314 return (__m512) 315 __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, 316 _mm512_undefined_ps (), 317 (__mmask16) -1); 318 } 319 320 extern __inline __m512 321 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 322 _mm512_mask_broadcast_f32x8 (__m512 __O, __mmask16 __M, __m256 __A) 323 { 324 return (__m512) __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, 325 (__v16sf)__O, 326 __M); 327 } 328 329 extern __inline __m512 330 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 331 _mm512_maskz_broadcast_f32x8 (__mmask16 __M, __m256 __A) 332 { 333 return (__m512) __builtin_ia32_broadcastf32x8_512_mask ((__v8sf) __A, 334 (__v16sf) 335 _mm512_setzero_ps (), 336 __M); 337 } 338 339 extern __inline __m512i 340 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 341 _mm512_broadcast_i32x8 (__m256i __A) 342 { 343 return (__m512i) 344 __builtin_ia32_broadcasti32x8_512_mask ((__v8si) __A, 345 (__v16si) 346 _mm512_undefined_epi32 (), 347 (__mmask16) -1); 348 } 349 350 extern __inline __m512i 351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 352 _mm512_mask_broadcast_i32x8 (__m512i __O, __mmask16 __M, __m256i __A) 353 { 354 return (__m512i) __builtin_ia32_broadcasti32x8_512_mask ((__v8si) 355 __A, 356 (__v16si)__O, 357 __M); 358 } 359 360 extern __inline __m512i 361 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 362 _mm512_maskz_broadcast_i32x8 (__mmask16 __M, __m256i __A) 363 { 364 return (__m512i) __builtin_ia32_broadcasti32x8_512_mask ((__v8si) 365 __A, 366 (__v16si) 367 _mm512_setzero_si512 (), 368 __M); 369 } 370 371 extern __inline __m512i 372 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 373 _mm512_mullo_epi64 (__m512i __A, __m512i __B) 374 { 375 return (__m512i) ((__v8du) __A * (__v8du) __B); 376 } 377 378 extern __inline __m512i 379 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 380 _mm512_mask_mullo_epi64 (__m512i __W, __mmask8 __U, __m512i __A, 381 __m512i __B) 382 { 383 return (__m512i) __builtin_ia32_pmullq512_mask ((__v8di) __A, 384 (__v8di) __B, 385 (__v8di) __W, 386 (__mmask8) __U); 387 } 388 389 extern __inline __m512i 390 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 391 _mm512_maskz_mullo_epi64 (__mmask8 __U, __m512i __A, __m512i __B) 392 { 393 return (__m512i) __builtin_ia32_pmullq512_mask ((__v8di) __A, 394 (__v8di) __B, 395 (__v8di) 396 _mm512_setzero_si512 (), 397 (__mmask8) __U); 398 } 399 400 extern __inline __m512d 401 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 402 _mm512_xor_pd (__m512d __A, __m512d __B) 403 { 404 return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, 405 (__v8df) __B, 406 (__v8df) 407 _mm512_setzero_pd (), 408 (__mmask8) -1); 409 } 410 411 extern __inline __m512d 412 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 413 _mm512_mask_xor_pd (__m512d __W, __mmask8 __U, __m512d __A, 414 __m512d __B) 415 { 416 return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, 417 (__v8df) __B, 418 (__v8df) __W, 419 (__mmask8) __U); 420 } 421 422 extern __inline __m512d 423 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 424 _mm512_maskz_xor_pd (__mmask8 __U, __m512d __A, __m512d __B) 425 { 426 return (__m512d) __builtin_ia32_xorpd512_mask ((__v8df) __A, 427 (__v8df) __B, 428 (__v8df) 429 _mm512_setzero_pd (), 430 (__mmask8) __U); 431 } 432 433 extern __inline __m512 434 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 435 _mm512_xor_ps (__m512 __A, __m512 __B) 436 { 437 return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, 438 (__v16sf) __B, 439 (__v16sf) 440 _mm512_setzero_ps (), 441 (__mmask16) -1); 442 } 443 444 extern __inline __m512 445 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 446 _mm512_mask_xor_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) 447 { 448 return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, 449 (__v16sf) __B, 450 (__v16sf) __W, 451 (__mmask16) __U); 452 } 453 454 extern __inline __m512 455 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 456 _mm512_maskz_xor_ps (__mmask16 __U, __m512 __A, __m512 __B) 457 { 458 return (__m512) __builtin_ia32_xorps512_mask ((__v16sf) __A, 459 (__v16sf) __B, 460 (__v16sf) 461 _mm512_setzero_ps (), 462 (__mmask16) __U); 463 } 464 465 extern __inline __m512d 466 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 467 _mm512_or_pd (__m512d __A, __m512d __B) 468 { 469 return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, 470 (__v8df) __B, 471 (__v8df) 472 _mm512_setzero_pd (), 473 (__mmask8) -1); 474 } 475 476 extern __inline __m512d 477 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 478 _mm512_mask_or_pd (__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) 479 { 480 return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, 481 (__v8df) __B, 482 (__v8df) __W, 483 (__mmask8) __U); 484 } 485 486 extern __inline __m512d 487 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 488 _mm512_maskz_or_pd (__mmask8 __U, __m512d __A, __m512d __B) 489 { 490 return (__m512d) __builtin_ia32_orpd512_mask ((__v8df) __A, 491 (__v8df) __B, 492 (__v8df) 493 _mm512_setzero_pd (), 494 (__mmask8) __U); 495 } 496 497 extern __inline __m512 498 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 499 _mm512_or_ps (__m512 __A, __m512 __B) 500 { 501 return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, 502 (__v16sf) __B, 503 (__v16sf) 504 _mm512_setzero_ps (), 505 (__mmask16) -1); 506 } 507 508 extern __inline __m512 509 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 510 _mm512_mask_or_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) 511 { 512 return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, 513 (__v16sf) __B, 514 (__v16sf) __W, 515 (__mmask16) __U); 516 } 517 518 extern __inline __m512 519 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 520 _mm512_maskz_or_ps (__mmask16 __U, __m512 __A, __m512 __B) 521 { 522 return (__m512) __builtin_ia32_orps512_mask ((__v16sf) __A, 523 (__v16sf) __B, 524 (__v16sf) 525 _mm512_setzero_ps (), 526 (__mmask16) __U); 527 } 528 529 extern __inline __m512d 530 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 531 _mm512_and_pd (__m512d __A, __m512d __B) 532 { 533 return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, 534 (__v8df) __B, 535 (__v8df) 536 _mm512_setzero_pd (), 537 (__mmask8) -1); 538 } 539 540 extern __inline __m512d 541 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 542 _mm512_mask_and_pd (__m512d __W, __mmask8 __U, __m512d __A, 543 __m512d __B) 544 { 545 return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, 546 (__v8df) __B, 547 (__v8df) __W, 548 (__mmask8) __U); 549 } 550 551 extern __inline __m512d 552 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 553 _mm512_maskz_and_pd (__mmask8 __U, __m512d __A, __m512d __B) 554 { 555 return (__m512d) __builtin_ia32_andpd512_mask ((__v8df) __A, 556 (__v8df) __B, 557 (__v8df) 558 _mm512_setzero_pd (), 559 (__mmask8) __U); 560 } 561 562 extern __inline __m512 563 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 564 _mm512_and_ps (__m512 __A, __m512 __B) 565 { 566 return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, 567 (__v16sf) __B, 568 (__v16sf) 569 _mm512_setzero_ps (), 570 (__mmask16) -1); 571 } 572 573 extern __inline __m512 574 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 575 _mm512_mask_and_ps (__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) 576 { 577 return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, 578 (__v16sf) __B, 579 (__v16sf) __W, 580 (__mmask16) __U); 581 } 582 583 extern __inline __m512 584 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 585 _mm512_maskz_and_ps (__mmask16 __U, __m512 __A, __m512 __B) 586 { 587 return (__m512) __builtin_ia32_andps512_mask ((__v16sf) __A, 588 (__v16sf) __B, 589 (__v16sf) 590 _mm512_setzero_ps (), 591 (__mmask16) __U); 592 } 593 594 extern __inline __m512d 595 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 596 _mm512_andnot_pd (__m512d __A, __m512d __B) 597 { 598 return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, 599 (__v8df) __B, 600 (__v8df) 601 _mm512_setzero_pd (), 602 (__mmask8) -1); 603 } 604 605 extern __inline __m512d 606 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 607 _mm512_mask_andnot_pd (__m512d __W, __mmask8 __U, __m512d __A, 608 __m512d __B) 609 { 610 return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, 611 (__v8df) __B, 612 (__v8df) __W, 613 (__mmask8) __U); 614 } 615 616 extern __inline __m512d 617 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 618 _mm512_maskz_andnot_pd (__mmask8 __U, __m512d __A, __m512d __B) 619 { 620 return (__m512d) __builtin_ia32_andnpd512_mask ((__v8df) __A, 621 (__v8df) __B, 622 (__v8df) 623 _mm512_setzero_pd (), 624 (__mmask8) __U); 625 } 626 627 extern __inline __m512 628 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 629 _mm512_andnot_ps (__m512 __A, __m512 __B) 630 { 631 return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, 632 (__v16sf) __B, 633 (__v16sf) 634 _mm512_setzero_ps (), 635 (__mmask16) -1); 636 } 637 638 extern __inline __m512 639 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 640 _mm512_mask_andnot_ps (__m512 __W, __mmask16 __U, __m512 __A, 641 __m512 __B) 642 { 643 return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, 644 (__v16sf) __B, 645 (__v16sf) __W, 646 (__mmask16) __U); 647 } 648 649 extern __inline __m512 650 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 651 _mm512_maskz_andnot_ps (__mmask16 __U, __m512 __A, __m512 __B) 652 { 653 return (__m512) __builtin_ia32_andnps512_mask ((__v16sf) __A, 654 (__v16sf) __B, 655 (__v16sf) 656 _mm512_setzero_ps (), 657 (__mmask16) __U); 658 } 659 660 extern __inline __mmask16 661 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 662 _mm512_movepi32_mask (__m512i __A) 663 { 664 return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A); 665 } 666 667 extern __inline __mmask8 668 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 669 _mm512_movepi64_mask (__m512i __A) 670 { 671 return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A); 672 } 673 674 extern __inline __m512i 675 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 676 _mm512_movm_epi32 (__mmask16 __A) 677 { 678 return (__m512i) __builtin_ia32_cvtmask2d512 (__A); 679 } 680 681 extern __inline __m512i 682 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 683 _mm512_movm_epi64 (__mmask8 __A) 684 { 685 return (__m512i) __builtin_ia32_cvtmask2q512 (__A); 686 } 687 688 extern __inline __m512i 689 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 690 _mm512_cvttpd_epi64 (__m512d __A) 691 { 692 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 693 (__v8di) 694 _mm512_setzero_si512 (), 695 (__mmask8) -1, 696 _MM_FROUND_CUR_DIRECTION); 697 } 698 699 extern __inline __m512i 700 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 701 _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) 702 { 703 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 704 (__v8di) __W, 705 (__mmask8) __U, 706 _MM_FROUND_CUR_DIRECTION); 707 } 708 709 extern __inline __m512i 710 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 711 _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) 712 { 713 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 714 (__v8di) 715 _mm512_setzero_si512 (), 716 (__mmask8) __U, 717 _MM_FROUND_CUR_DIRECTION); 718 } 719 720 extern __inline __m512i 721 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 722 _mm512_cvttpd_epu64 (__m512d __A) 723 { 724 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 725 (__v8di) 726 _mm512_setzero_si512 (), 727 (__mmask8) -1, 728 _MM_FROUND_CUR_DIRECTION); 729 } 730 731 extern __inline __m512i 732 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 733 _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) 734 { 735 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 736 (__v8di) __W, 737 (__mmask8) __U, 738 _MM_FROUND_CUR_DIRECTION); 739 } 740 741 extern __inline __m512i 742 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 743 _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) 744 { 745 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 746 (__v8di) 747 _mm512_setzero_si512 (), 748 (__mmask8) __U, 749 _MM_FROUND_CUR_DIRECTION); 750 } 751 752 extern __inline __m512i 753 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 754 _mm512_cvttps_epi64 (__m256 __A) 755 { 756 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 757 (__v8di) 758 _mm512_setzero_si512 (), 759 (__mmask8) -1, 760 _MM_FROUND_CUR_DIRECTION); 761 } 762 763 extern __inline __m512i 764 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 765 _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) 766 { 767 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 768 (__v8di) __W, 769 (__mmask8) __U, 770 _MM_FROUND_CUR_DIRECTION); 771 } 772 773 extern __inline __m512i 774 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 775 _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) 776 { 777 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 778 (__v8di) 779 _mm512_setzero_si512 (), 780 (__mmask8) __U, 781 _MM_FROUND_CUR_DIRECTION); 782 } 783 784 extern __inline __m512i 785 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 786 _mm512_cvttps_epu64 (__m256 __A) 787 { 788 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 789 (__v8di) 790 _mm512_setzero_si512 (), 791 (__mmask8) -1, 792 _MM_FROUND_CUR_DIRECTION); 793 } 794 795 extern __inline __m512i 796 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 797 _mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) 798 { 799 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 800 (__v8di) __W, 801 (__mmask8) __U, 802 _MM_FROUND_CUR_DIRECTION); 803 } 804 805 extern __inline __m512i 806 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 807 _mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) 808 { 809 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 810 (__v8di) 811 _mm512_setzero_si512 (), 812 (__mmask8) __U, 813 _MM_FROUND_CUR_DIRECTION); 814 } 815 816 extern __inline __m512i 817 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 818 _mm512_cvtpd_epi64 (__m512d __A) 819 { 820 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 821 (__v8di) 822 _mm512_setzero_si512 (), 823 (__mmask8) -1, 824 _MM_FROUND_CUR_DIRECTION); 825 } 826 827 extern __inline __m512i 828 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 829 _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) 830 { 831 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 832 (__v8di) __W, 833 (__mmask8) __U, 834 _MM_FROUND_CUR_DIRECTION); 835 } 836 837 extern __inline __m512i 838 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 839 _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) 840 { 841 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 842 (__v8di) 843 _mm512_setzero_si512 (), 844 (__mmask8) __U, 845 _MM_FROUND_CUR_DIRECTION); 846 } 847 848 extern __inline __m512i 849 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 850 _mm512_cvtpd_epu64 (__m512d __A) 851 { 852 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 853 (__v8di) 854 _mm512_setzero_si512 (), 855 (__mmask8) -1, 856 _MM_FROUND_CUR_DIRECTION); 857 } 858 859 extern __inline __m512i 860 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 861 _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) 862 { 863 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 864 (__v8di) __W, 865 (__mmask8) __U, 866 _MM_FROUND_CUR_DIRECTION); 867 } 868 869 extern __inline __m512i 870 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 871 _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) 872 { 873 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 874 (__v8di) 875 _mm512_setzero_si512 (), 876 (__mmask8) __U, 877 _MM_FROUND_CUR_DIRECTION); 878 } 879 880 extern __inline __m512i 881 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 882 _mm512_cvtps_epi64 (__m256 __A) 883 { 884 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 885 (__v8di) 886 _mm512_setzero_si512 (), 887 (__mmask8) -1, 888 _MM_FROUND_CUR_DIRECTION); 889 } 890 891 extern __inline __m512i 892 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 893 _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) 894 { 895 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 896 (__v8di) __W, 897 (__mmask8) __U, 898 _MM_FROUND_CUR_DIRECTION); 899 } 900 901 extern __inline __m512i 902 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 903 _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) 904 { 905 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 906 (__v8di) 907 _mm512_setzero_si512 (), 908 (__mmask8) __U, 909 _MM_FROUND_CUR_DIRECTION); 910 } 911 912 extern __inline __m512i 913 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 914 _mm512_cvtps_epu64 (__m256 __A) 915 { 916 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 917 (__v8di) 918 _mm512_setzero_si512 (), 919 (__mmask8) -1, 920 _MM_FROUND_CUR_DIRECTION); 921 } 922 923 extern __inline __m512i 924 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 925 _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) 926 { 927 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 928 (__v8di) __W, 929 (__mmask8) __U, 930 _MM_FROUND_CUR_DIRECTION); 931 } 932 933 extern __inline __m512i 934 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 935 _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) 936 { 937 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 938 (__v8di) 939 _mm512_setzero_si512 (), 940 (__mmask8) __U, 941 _MM_FROUND_CUR_DIRECTION); 942 } 943 944 extern __inline __m256 945 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 946 _mm512_cvtepi64_ps (__m512i __A) 947 { 948 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 949 (__v8sf) 950 _mm256_setzero_ps (), 951 (__mmask8) -1, 952 _MM_FROUND_CUR_DIRECTION); 953 } 954 955 extern __inline __m256 956 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 957 _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) 958 { 959 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 960 (__v8sf) __W, 961 (__mmask8) __U, 962 _MM_FROUND_CUR_DIRECTION); 963 } 964 965 extern __inline __m256 966 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 967 _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) 968 { 969 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 970 (__v8sf) 971 _mm256_setzero_ps (), 972 (__mmask8) __U, 973 _MM_FROUND_CUR_DIRECTION); 974 } 975 976 extern __inline __m256 977 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 978 _mm512_cvtepu64_ps (__m512i __A) 979 { 980 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 981 (__v8sf) 982 _mm256_setzero_ps (), 983 (__mmask8) -1, 984 _MM_FROUND_CUR_DIRECTION); 985 } 986 987 extern __inline __m256 988 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 989 _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) 990 { 991 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 992 (__v8sf) __W, 993 (__mmask8) __U, 994 _MM_FROUND_CUR_DIRECTION); 995 } 996 997 extern __inline __m256 998 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 999 _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) 1000 { 1001 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1002 (__v8sf) 1003 _mm256_setzero_ps (), 1004 (__mmask8) __U, 1005 _MM_FROUND_CUR_DIRECTION); 1006 } 1007 1008 extern __inline __m512d 1009 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1010 _mm512_cvtepi64_pd (__m512i __A) 1011 { 1012 return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1013 (__v8df) 1014 _mm512_setzero_pd (), 1015 (__mmask8) -1, 1016 _MM_FROUND_CUR_DIRECTION); 1017 } 1018 1019 extern __inline __m512d 1020 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1021 _mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) 1022 { 1023 return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1024 (__v8df) __W, 1025 (__mmask8) __U, 1026 _MM_FROUND_CUR_DIRECTION); 1027 } 1028 1029 extern __inline __m512d 1030 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1031 _mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) 1032 { 1033 return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1034 (__v8df) 1035 _mm512_setzero_pd (), 1036 (__mmask8) __U, 1037 _MM_FROUND_CUR_DIRECTION); 1038 } 1039 1040 extern __inline __m512d 1041 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1042 _mm512_cvtepu64_pd (__m512i __A) 1043 { 1044 return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1045 (__v8df) 1046 _mm512_setzero_pd (), 1047 (__mmask8) -1, 1048 _MM_FROUND_CUR_DIRECTION); 1049 } 1050 1051 extern __inline __m512d 1052 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1053 _mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) 1054 { 1055 return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1056 (__v8df) __W, 1057 (__mmask8) __U, 1058 _MM_FROUND_CUR_DIRECTION); 1059 } 1060 1061 extern __inline __m512d 1062 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1063 _mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) 1064 { 1065 return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1066 (__v8df) 1067 _mm512_setzero_pd (), 1068 (__mmask8) __U, 1069 _MM_FROUND_CUR_DIRECTION); 1070 } 1071 1072 #ifdef __OPTIMIZE__ 1073 extern __inline __mmask8 1074 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1075 _kshiftli_mask8 (__mmask8 __A, unsigned int __B) 1076 { 1077 return (__mmask8) __builtin_ia32_kshiftliqi ((__mmask8) __A, (__mmask8) __B); 1078 } 1079 1080 extern __inline __mmask8 1081 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1082 _kshiftri_mask8 (__mmask8 __A, unsigned int __B) 1083 { 1084 return (__mmask8) __builtin_ia32_kshiftriqi ((__mmask8) __A, (__mmask8) __B); 1085 } 1086 1087 extern __inline __m512d 1088 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1089 _mm512_range_pd (__m512d __A, __m512d __B, int __C) 1090 { 1091 return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1092 (__v8df) __B, __C, 1093 (__v8df) 1094 _mm512_setzero_pd (), 1095 (__mmask8) -1, 1096 _MM_FROUND_CUR_DIRECTION); 1097 } 1098 1099 extern __inline __m512d 1100 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1101 _mm512_mask_range_pd (__m512d __W, __mmask8 __U, 1102 __m512d __A, __m512d __B, int __C) 1103 { 1104 return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1105 (__v8df) __B, __C, 1106 (__v8df) __W, 1107 (__mmask8) __U, 1108 _MM_FROUND_CUR_DIRECTION); 1109 } 1110 1111 extern __inline __m512d 1112 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1113 _mm512_maskz_range_pd (__mmask8 __U, __m512d __A, __m512d __B, int __C) 1114 { 1115 return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1116 (__v8df) __B, __C, 1117 (__v8df) 1118 _mm512_setzero_pd (), 1119 (__mmask8) __U, 1120 _MM_FROUND_CUR_DIRECTION); 1121 } 1122 1123 extern __inline __m512 1124 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1125 _mm512_range_ps (__m512 __A, __m512 __B, int __C) 1126 { 1127 return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1128 (__v16sf) __B, __C, 1129 (__v16sf) 1130 _mm512_setzero_ps (), 1131 (__mmask16) -1, 1132 _MM_FROUND_CUR_DIRECTION); 1133 } 1134 1135 extern __inline __m512 1136 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1137 _mm512_mask_range_ps (__m512 __W, __mmask16 __U, 1138 __m512 __A, __m512 __B, int __C) 1139 { 1140 return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1141 (__v16sf) __B, __C, 1142 (__v16sf) __W, 1143 (__mmask16) __U, 1144 _MM_FROUND_CUR_DIRECTION); 1145 } 1146 1147 extern __inline __m512 1148 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1149 _mm512_maskz_range_ps (__mmask16 __U, __m512 __A, __m512 __B, int __C) 1150 { 1151 return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1152 (__v16sf) __B, __C, 1153 (__v16sf) 1154 _mm512_setzero_ps (), 1155 (__mmask16) __U, 1156 _MM_FROUND_CUR_DIRECTION); 1157 } 1158 1159 extern __inline __m128d 1160 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1161 _mm_reduce_sd (__m128d __A, __m128d __B, int __C) 1162 { 1163 return (__m128d) __builtin_ia32_reducesd ((__v2df) __A, 1164 (__v2df) __B, __C); 1165 } 1166 1167 extern __inline __m128 1168 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1169 _mm_reduce_ss (__m128 __A, __m128 __B, int __C) 1170 { 1171 return (__m128) __builtin_ia32_reducess ((__v4sf) __A, 1172 (__v4sf) __B, __C); 1173 } 1174 1175 extern __inline __m128d 1176 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1177 _mm_range_sd (__m128d __A, __m128d __B, int __C) 1178 { 1179 return (__m128d) __builtin_ia32_rangesd128_round ((__v2df) __A, 1180 (__v2df) __B, __C, 1181 _MM_FROUND_CUR_DIRECTION); 1182 } 1183 1184 1185 extern __inline __m128 1186 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1187 _mm_range_ss (__m128 __A, __m128 __B, int __C) 1188 { 1189 return (__m128) __builtin_ia32_rangess128_round ((__v4sf) __A, 1190 (__v4sf) __B, __C, 1191 _MM_FROUND_CUR_DIRECTION); 1192 } 1193 1194 extern __inline __m128d 1195 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1196 _mm_range_round_sd (__m128d __A, __m128d __B, int __C, const int __R) 1197 { 1198 return (__m128d) __builtin_ia32_rangesd128_round ((__v2df) __A, 1199 (__v2df) __B, __C, 1200 __R); 1201 } 1202 1203 extern __inline __m128 1204 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1205 _mm_range_round_ss (__m128 __A, __m128 __B, int __C, const int __R) 1206 { 1207 return (__m128) __builtin_ia32_rangess128_round ((__v4sf) __A, 1208 (__v4sf) __B, __C, 1209 __R); 1210 } 1211 1212 extern __inline __mmask8 1213 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1214 _mm_fpclass_ss_mask (__m128 __A, const int __imm) 1215 { 1216 return (__mmask8) __builtin_ia32_fpclassss ((__v4sf) __A, __imm); 1217 } 1218 1219 extern __inline __mmask8 1220 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1221 _mm_fpclass_sd_mask (__m128d __A, const int __imm) 1222 { 1223 return (__mmask8) __builtin_ia32_fpclasssd ((__v2df) __A, __imm); 1224 } 1225 1226 extern __inline __m512i 1227 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1228 _mm512_cvtt_roundpd_epi64 (__m512d __A, const int __R) 1229 { 1230 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 1231 (__v8di) 1232 _mm512_setzero_si512 (), 1233 (__mmask8) -1, 1234 __R); 1235 } 1236 1237 extern __inline __m512i 1238 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1239 _mm512_mask_cvtt_roundpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A, 1240 const int __R) 1241 { 1242 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 1243 (__v8di) __W, 1244 (__mmask8) __U, 1245 __R); 1246 } 1247 1248 extern __inline __m512i 1249 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1250 _mm512_maskz_cvtt_roundpd_epi64 (__mmask8 __U, __m512d __A, 1251 const int __R) 1252 { 1253 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, 1254 (__v8di) 1255 _mm512_setzero_si512 (), 1256 (__mmask8) __U, 1257 __R); 1258 } 1259 1260 extern __inline __m512i 1261 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1262 _mm512_cvtt_roundpd_epu64 (__m512d __A, const int __R) 1263 { 1264 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 1265 (__v8di) 1266 _mm512_setzero_si512 (), 1267 (__mmask8) -1, 1268 __R); 1269 } 1270 1271 extern __inline __m512i 1272 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1273 _mm512_mask_cvtt_roundpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A, 1274 const int __R) 1275 { 1276 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 1277 (__v8di) __W, 1278 (__mmask8) __U, 1279 __R); 1280 } 1281 1282 extern __inline __m512i 1283 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1284 _mm512_maskz_cvtt_roundpd_epu64 (__mmask8 __U, __m512d __A, 1285 const int __R) 1286 { 1287 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, 1288 (__v8di) 1289 _mm512_setzero_si512 (), 1290 (__mmask8) __U, 1291 __R); 1292 } 1293 1294 extern __inline __m512i 1295 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1296 _mm512_cvtt_roundps_epi64 (__m256 __A, const int __R) 1297 { 1298 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 1299 (__v8di) 1300 _mm512_setzero_si512 (), 1301 (__mmask8) -1, 1302 __R); 1303 } 1304 1305 extern __inline __m512i 1306 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1307 _mm512_mask_cvtt_roundps_epi64 (__m512i __W, __mmask8 __U, __m256 __A, 1308 const int __R) 1309 { 1310 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 1311 (__v8di) __W, 1312 (__mmask8) __U, 1313 __R); 1314 } 1315 1316 extern __inline __m512i 1317 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1318 _mm512_maskz_cvtt_roundps_epi64 (__mmask8 __U, __m256 __A, 1319 const int __R) 1320 { 1321 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, 1322 (__v8di) 1323 _mm512_setzero_si512 (), 1324 (__mmask8) __U, 1325 __R); 1326 } 1327 1328 extern __inline __m512i 1329 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1330 _mm512_cvtt_roundps_epu64 (__m256 __A, const int __R) 1331 { 1332 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 1333 (__v8di) 1334 _mm512_setzero_si512 (), 1335 (__mmask8) -1, 1336 __R); 1337 } 1338 1339 extern __inline __m512i 1340 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1341 _mm512_mask_cvtt_roundps_epu64 (__m512i __W, __mmask8 __U, __m256 __A, 1342 const int __R) 1343 { 1344 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 1345 (__v8di) __W, 1346 (__mmask8) __U, 1347 __R); 1348 } 1349 1350 extern __inline __m512i 1351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1352 _mm512_maskz_cvtt_roundps_epu64 (__mmask8 __U, __m256 __A, 1353 const int __R) 1354 { 1355 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, 1356 (__v8di) 1357 _mm512_setzero_si512 (), 1358 (__mmask8) __U, 1359 __R); 1360 } 1361 1362 extern __inline __m512i 1363 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1364 _mm512_cvt_roundpd_epi64 (__m512d __A, const int __R) 1365 { 1366 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 1367 (__v8di) 1368 _mm512_setzero_si512 (), 1369 (__mmask8) -1, 1370 __R); 1371 } 1372 1373 extern __inline __m512i 1374 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1375 _mm512_mask_cvt_roundpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A, 1376 const int __R) 1377 { 1378 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 1379 (__v8di) __W, 1380 (__mmask8) __U, 1381 __R); 1382 } 1383 1384 extern __inline __m512i 1385 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1386 _mm512_maskz_cvt_roundpd_epi64 (__mmask8 __U, __m512d __A, 1387 const int __R) 1388 { 1389 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, 1390 (__v8di) 1391 _mm512_setzero_si512 (), 1392 (__mmask8) __U, 1393 __R); 1394 } 1395 1396 extern __inline __m512i 1397 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1398 _mm512_cvt_roundpd_epu64 (__m512d __A, const int __R) 1399 { 1400 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 1401 (__v8di) 1402 _mm512_setzero_si512 (), 1403 (__mmask8) -1, 1404 __R); 1405 } 1406 1407 extern __inline __m512i 1408 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1409 _mm512_mask_cvt_roundpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A, 1410 const int __R) 1411 { 1412 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 1413 (__v8di) __W, 1414 (__mmask8) __U, 1415 __R); 1416 } 1417 1418 extern __inline __m512i 1419 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1420 _mm512_maskz_cvt_roundpd_epu64 (__mmask8 __U, __m512d __A, 1421 const int __R) 1422 { 1423 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, 1424 (__v8di) 1425 _mm512_setzero_si512 (), 1426 (__mmask8) __U, 1427 __R); 1428 } 1429 1430 extern __inline __m512i 1431 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1432 _mm512_cvt_roundps_epi64 (__m256 __A, const int __R) 1433 { 1434 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 1435 (__v8di) 1436 _mm512_setzero_si512 (), 1437 (__mmask8) -1, 1438 __R); 1439 } 1440 1441 extern __inline __m512i 1442 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1443 _mm512_mask_cvt_roundps_epi64 (__m512i __W, __mmask8 __U, __m256 __A, 1444 const int __R) 1445 { 1446 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 1447 (__v8di) __W, 1448 (__mmask8) __U, 1449 __R); 1450 } 1451 1452 extern __inline __m512i 1453 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1454 _mm512_maskz_cvt_roundps_epi64 (__mmask8 __U, __m256 __A, 1455 const int __R) 1456 { 1457 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, 1458 (__v8di) 1459 _mm512_setzero_si512 (), 1460 (__mmask8) __U, 1461 __R); 1462 } 1463 1464 extern __inline __m512i 1465 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1466 _mm512_cvt_roundps_epu64 (__m256 __A, const int __R) 1467 { 1468 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 1469 (__v8di) 1470 _mm512_setzero_si512 (), 1471 (__mmask8) -1, 1472 __R); 1473 } 1474 1475 extern __inline __m512i 1476 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1477 _mm512_mask_cvt_roundps_epu64 (__m512i __W, __mmask8 __U, __m256 __A, 1478 const int __R) 1479 { 1480 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 1481 (__v8di) __W, 1482 (__mmask8) __U, 1483 __R); 1484 } 1485 1486 extern __inline __m512i 1487 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1488 _mm512_maskz_cvt_roundps_epu64 (__mmask8 __U, __m256 __A, 1489 const int __R) 1490 { 1491 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, 1492 (__v8di) 1493 _mm512_setzero_si512 (), 1494 (__mmask8) __U, 1495 __R); 1496 } 1497 1498 extern __inline __m256 1499 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1500 _mm512_cvt_roundepi64_ps (__m512i __A, const int __R) 1501 { 1502 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 1503 (__v8sf) 1504 _mm256_setzero_ps (), 1505 (__mmask8) -1, 1506 __R); 1507 } 1508 1509 extern __inline __m256 1510 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1511 _mm512_mask_cvt_roundepi64_ps (__m256 __W, __mmask8 __U, __m512i __A, 1512 const int __R) 1513 { 1514 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 1515 (__v8sf) __W, 1516 (__mmask8) __U, 1517 __R); 1518 } 1519 1520 extern __inline __m256 1521 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1522 _mm512_maskz_cvt_roundepi64_ps (__mmask8 __U, __m512i __A, 1523 const int __R) 1524 { 1525 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, 1526 (__v8sf) 1527 _mm256_setzero_ps (), 1528 (__mmask8) __U, 1529 __R); 1530 } 1531 1532 extern __inline __m256 1533 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1534 _mm512_cvt_roundepu64_ps (__m512i __A, const int __R) 1535 { 1536 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1537 (__v8sf) 1538 _mm256_setzero_ps (), 1539 (__mmask8) -1, 1540 __R); 1541 } 1542 1543 extern __inline __m256 1544 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1545 _mm512_mask_cvt_roundepu64_ps (__m256 __W, __mmask8 __U, __m512i __A, 1546 const int __R) 1547 { 1548 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1549 (__v8sf) __W, 1550 (__mmask8) __U, 1551 __R); 1552 } 1553 1554 extern __inline __m256 1555 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1556 _mm512_maskz_cvt_roundepu64_ps (__mmask8 __U, __m512i __A, 1557 const int __R) 1558 { 1559 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, 1560 (__v8sf) 1561 _mm256_setzero_ps (), 1562 (__mmask8) __U, 1563 __R); 1564 } 1565 1566 extern __inline __m512d 1567 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1568 _mm512_cvt_roundepi64_pd (__m512i __A, const int __R) 1569 { 1570 return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1571 (__v8df) 1572 _mm512_setzero_pd (), 1573 (__mmask8) -1, 1574 __R); 1575 } 1576 1577 extern __inline __m512d 1578 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1579 _mm512_mask_cvt_roundepi64_pd (__m512d __W, __mmask8 __U, __m512i __A, 1580 const int __R) 1581 { 1582 return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1583 (__v8df) __W, 1584 (__mmask8) __U, 1585 __R); 1586 } 1587 1588 extern __inline __m512d 1589 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1590 _mm512_maskz_cvt_roundepi64_pd (__mmask8 __U, __m512i __A, 1591 const int __R) 1592 { 1593 return (__m512d) __builtin_ia32_cvtqq2pd512_mask ((__v8di) __A, 1594 (__v8df) 1595 _mm512_setzero_pd (), 1596 (__mmask8) __U, 1597 __R); 1598 } 1599 1600 extern __inline __m512d 1601 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1602 _mm512_cvt_roundepu64_pd (__m512i __A, const int __R) 1603 { 1604 return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1605 (__v8df) 1606 _mm512_setzero_pd (), 1607 (__mmask8) -1, 1608 __R); 1609 } 1610 1611 extern __inline __m512d 1612 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1613 _mm512_mask_cvt_roundepu64_pd (__m512d __W, __mmask8 __U, __m512i __A, 1614 const int __R) 1615 { 1616 return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1617 (__v8df) __W, 1618 (__mmask8) __U, 1619 __R); 1620 } 1621 1622 extern __inline __m512d 1623 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1624 _mm512_maskz_cvt_roundepu64_pd (__mmask8 __U, __m512i __A, 1625 const int __R) 1626 { 1627 return (__m512d) __builtin_ia32_cvtuqq2pd512_mask ((__v8di) __A, 1628 (__v8df) 1629 _mm512_setzero_pd (), 1630 (__mmask8) __U, 1631 __R); 1632 } 1633 1634 extern __inline __m512d 1635 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1636 _mm512_reduce_pd (__m512d __A, int __B) 1637 { 1638 return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, 1639 (__v8df) 1640 _mm512_setzero_pd (), 1641 (__mmask8) -1); 1642 } 1643 1644 extern __inline __m512d 1645 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1646 _mm512_mask_reduce_pd (__m512d __W, __mmask8 __U, __m512d __A, int __B) 1647 { 1648 return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, 1649 (__v8df) __W, 1650 (__mmask8) __U); 1651 } 1652 1653 extern __inline __m512d 1654 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1655 _mm512_maskz_reduce_pd (__mmask8 __U, __m512d __A, int __B) 1656 { 1657 return (__m512d) __builtin_ia32_reducepd512_mask ((__v8df) __A, __B, 1658 (__v8df) 1659 _mm512_setzero_pd (), 1660 (__mmask8) __U); 1661 } 1662 1663 extern __inline __m512 1664 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1665 _mm512_reduce_ps (__m512 __A, int __B) 1666 { 1667 return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, 1668 (__v16sf) 1669 _mm512_setzero_ps (), 1670 (__mmask16) -1); 1671 } 1672 1673 extern __inline __m512 1674 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1675 _mm512_mask_reduce_ps (__m512 __W, __mmask16 __U, __m512 __A, int __B) 1676 { 1677 return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, 1678 (__v16sf) __W, 1679 (__mmask16) __U); 1680 } 1681 1682 extern __inline __m512 1683 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1684 _mm512_maskz_reduce_ps (__mmask16 __U, __m512 __A, int __B) 1685 { 1686 return (__m512) __builtin_ia32_reduceps512_mask ((__v16sf) __A, __B, 1687 (__v16sf) 1688 _mm512_setzero_ps (), 1689 (__mmask16) __U); 1690 } 1691 1692 extern __inline __m256 1693 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1694 _mm512_extractf32x8_ps (__m512 __A, const int __imm) 1695 { 1696 return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, 1697 __imm, 1698 (__v8sf) 1699 _mm256_setzero_ps (), 1700 (__mmask8) -1); 1701 } 1702 1703 extern __inline __m256 1704 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1705 _mm512_mask_extractf32x8_ps (__m256 __W, __mmask8 __U, __m512 __A, 1706 const int __imm) 1707 { 1708 return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, 1709 __imm, 1710 (__v8sf) __W, 1711 (__mmask8) __U); 1712 } 1713 1714 extern __inline __m256 1715 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1716 _mm512_maskz_extractf32x8_ps (__mmask8 __U, __m512 __A, 1717 const int __imm) 1718 { 1719 return (__m256) __builtin_ia32_extractf32x8_mask ((__v16sf) __A, 1720 __imm, 1721 (__v8sf) 1722 _mm256_setzero_ps (), 1723 (__mmask8) __U); 1724 } 1725 1726 extern __inline __m128d 1727 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1728 _mm512_extractf64x2_pd (__m512d __A, const int __imm) 1729 { 1730 return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, 1731 __imm, 1732 (__v2df) 1733 _mm_setzero_pd (), 1734 (__mmask8) -1); 1735 } 1736 1737 extern __inline __m128d 1738 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1739 _mm512_mask_extractf64x2_pd (__m128d __W, __mmask8 __U, __m512d __A, 1740 const int __imm) 1741 { 1742 return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, 1743 __imm, 1744 (__v2df) __W, 1745 (__mmask8) 1746 __U); 1747 } 1748 1749 extern __inline __m128d 1750 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1751 _mm512_maskz_extractf64x2_pd (__mmask8 __U, __m512d __A, 1752 const int __imm) 1753 { 1754 return (__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df) __A, 1755 __imm, 1756 (__v2df) 1757 _mm_setzero_pd (), 1758 (__mmask8) 1759 __U); 1760 } 1761 1762 extern __inline __m256i 1763 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1764 _mm512_extracti32x8_epi32 (__m512i __A, const int __imm) 1765 { 1766 return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, 1767 __imm, 1768 (__v8si) 1769 _mm256_setzero_si256 (), 1770 (__mmask8) -1); 1771 } 1772 1773 extern __inline __m256i 1774 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1775 _mm512_mask_extracti32x8_epi32 (__m256i __W, __mmask8 __U, __m512i __A, 1776 const int __imm) 1777 { 1778 return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, 1779 __imm, 1780 (__v8si) __W, 1781 (__mmask8) __U); 1782 } 1783 1784 extern __inline __m256i 1785 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1786 _mm512_maskz_extracti32x8_epi32 (__mmask8 __U, __m512i __A, 1787 const int __imm) 1788 { 1789 return (__m256i) __builtin_ia32_extracti32x8_mask ((__v16si) __A, 1790 __imm, 1791 (__v8si) 1792 _mm256_setzero_si256 (), 1793 (__mmask8) __U); 1794 } 1795 1796 extern __inline __m128i 1797 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1798 _mm512_extracti64x2_epi64 (__m512i __A, const int __imm) 1799 { 1800 return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, 1801 __imm, 1802 (__v2di) 1803 _mm_setzero_si128 (), 1804 (__mmask8) -1); 1805 } 1806 1807 extern __inline __m128i 1808 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1809 _mm512_mask_extracti64x2_epi64 (__m128i __W, __mmask8 __U, __m512i __A, 1810 const int __imm) 1811 { 1812 return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, 1813 __imm, 1814 (__v2di) __W, 1815 (__mmask8) 1816 __U); 1817 } 1818 1819 extern __inline __m128i 1820 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1821 _mm512_maskz_extracti64x2_epi64 (__mmask8 __U, __m512i __A, 1822 const int __imm) 1823 { 1824 return (__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di) __A, 1825 __imm, 1826 (__v2di) 1827 _mm_setzero_si128 (), 1828 (__mmask8) 1829 __U); 1830 } 1831 1832 extern __inline __m512d 1833 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1834 _mm512_range_round_pd (__m512d __A, __m512d __B, int __C, 1835 const int __R) 1836 { 1837 return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1838 (__v8df) __B, __C, 1839 (__v8df) 1840 _mm512_setzero_pd (), 1841 (__mmask8) -1, 1842 __R); 1843 } 1844 1845 extern __inline __m512d 1846 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1847 _mm512_mask_range_round_pd (__m512d __W, __mmask8 __U, 1848 __m512d __A, __m512d __B, int __C, 1849 const int __R) 1850 { 1851 return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1852 (__v8df) __B, __C, 1853 (__v8df) __W, 1854 (__mmask8) __U, 1855 __R); 1856 } 1857 1858 extern __inline __m512d 1859 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1860 _mm512_maskz_range_round_pd (__mmask8 __U, __m512d __A, __m512d __B, 1861 int __C, const int __R) 1862 { 1863 return (__m512d) __builtin_ia32_rangepd512_mask ((__v8df) __A, 1864 (__v8df) __B, __C, 1865 (__v8df) 1866 _mm512_setzero_pd (), 1867 (__mmask8) __U, 1868 __R); 1869 } 1870 1871 extern __inline __m512 1872 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1873 _mm512_range_round_ps (__m512 __A, __m512 __B, int __C, const int __R) 1874 { 1875 return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1876 (__v16sf) __B, __C, 1877 (__v16sf) 1878 _mm512_setzero_ps (), 1879 (__mmask16) -1, 1880 __R); 1881 } 1882 1883 extern __inline __m512 1884 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1885 _mm512_mask_range_round_ps (__m512 __W, __mmask16 __U, 1886 __m512 __A, __m512 __B, int __C, 1887 const int __R) 1888 { 1889 return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1890 (__v16sf) __B, __C, 1891 (__v16sf) __W, 1892 (__mmask16) __U, 1893 __R); 1894 } 1895 1896 extern __inline __m512 1897 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1898 _mm512_maskz_range_round_ps (__mmask16 __U, __m512 __A, __m512 __B, 1899 int __C, const int __R) 1900 { 1901 return (__m512) __builtin_ia32_rangeps512_mask ((__v16sf) __A, 1902 (__v16sf) __B, __C, 1903 (__v16sf) 1904 _mm512_setzero_ps (), 1905 (__mmask16) __U, 1906 __R); 1907 } 1908 1909 extern __inline __m512i 1910 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1911 _mm512_inserti32x8 (__m512i __A, __m256i __B, const int __imm) 1912 { 1913 return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, 1914 (__v8si) __B, 1915 __imm, 1916 (__v16si) 1917 _mm512_setzero_si512 (), 1918 (__mmask16) -1); 1919 } 1920 1921 extern __inline __m512i 1922 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1923 _mm512_mask_inserti32x8 (__m512i __W, __mmask16 __U, __m512i __A, 1924 __m256i __B, const int __imm) 1925 { 1926 return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, 1927 (__v8si) __B, 1928 __imm, 1929 (__v16si) __W, 1930 (__mmask16) __U); 1931 } 1932 1933 extern __inline __m512i 1934 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1935 _mm512_maskz_inserti32x8 (__mmask16 __U, __m512i __A, __m256i __B, 1936 const int __imm) 1937 { 1938 return (__m512i) __builtin_ia32_inserti32x8_mask ((__v16si) __A, 1939 (__v8si) __B, 1940 __imm, 1941 (__v16si) 1942 _mm512_setzero_si512 (), 1943 (__mmask16) __U); 1944 } 1945 1946 extern __inline __m512 1947 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1948 _mm512_insertf32x8 (__m512 __A, __m256 __B, const int __imm) 1949 { 1950 return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, 1951 (__v8sf) __B, 1952 __imm, 1953 (__v16sf) 1954 _mm512_setzero_ps (), 1955 (__mmask16) -1); 1956 } 1957 1958 extern __inline __m512 1959 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1960 _mm512_mask_insertf32x8 (__m512 __W, __mmask16 __U, __m512 __A, 1961 __m256 __B, const int __imm) 1962 { 1963 return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, 1964 (__v8sf) __B, 1965 __imm, 1966 (__v16sf) __W, 1967 (__mmask16) __U); 1968 } 1969 1970 extern __inline __m512 1971 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1972 _mm512_maskz_insertf32x8 (__mmask16 __U, __m512 __A, __m256 __B, 1973 const int __imm) 1974 { 1975 return (__m512) __builtin_ia32_insertf32x8_mask ((__v16sf) __A, 1976 (__v8sf) __B, 1977 __imm, 1978 (__v16sf) 1979 _mm512_setzero_ps (), 1980 (__mmask16) __U); 1981 } 1982 1983 extern __inline __m512i 1984 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1985 _mm512_inserti64x2 (__m512i __A, __m128i __B, const int __imm) 1986 { 1987 return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, 1988 (__v2di) __B, 1989 __imm, 1990 (__v8di) 1991 _mm512_setzero_si512 (), 1992 (__mmask8) -1); 1993 } 1994 1995 extern __inline __m512i 1996 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 1997 _mm512_mask_inserti64x2 (__m512i __W, __mmask8 __U, __m512i __A, 1998 __m128i __B, const int __imm) 1999 { 2000 return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, 2001 (__v2di) __B, 2002 __imm, 2003 (__v8di) __W, 2004 (__mmask8) 2005 __U); 2006 } 2007 2008 extern __inline __m512i 2009 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2010 _mm512_maskz_inserti64x2 (__mmask8 __U, __m512i __A, __m128i __B, 2011 const int __imm) 2012 { 2013 return (__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di) __A, 2014 (__v2di) __B, 2015 __imm, 2016 (__v8di) 2017 _mm512_setzero_si512 (), 2018 (__mmask8) 2019 __U); 2020 } 2021 2022 extern __inline __m512d 2023 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2024 _mm512_insertf64x2 (__m512d __A, __m128d __B, const int __imm) 2025 { 2026 return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, 2027 (__v2df) __B, 2028 __imm, 2029 (__v8df) 2030 _mm512_setzero_pd (), 2031 (__mmask8) -1); 2032 } 2033 2034 extern __inline __m512d 2035 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2036 _mm512_mask_insertf64x2 (__m512d __W, __mmask8 __U, __m512d __A, 2037 __m128d __B, const int __imm) 2038 { 2039 return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, 2040 (__v2df) __B, 2041 __imm, 2042 (__v8df) __W, 2043 (__mmask8) 2044 __U); 2045 } 2046 2047 extern __inline __m512d 2048 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2049 _mm512_maskz_insertf64x2 (__mmask8 __U, __m512d __A, __m128d __B, 2050 const int __imm) 2051 { 2052 return (__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df) __A, 2053 (__v2df) __B, 2054 __imm, 2055 (__v8df) 2056 _mm512_setzero_pd (), 2057 (__mmask8) 2058 __U); 2059 } 2060 2061 extern __inline __mmask8 2062 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2063 _mm512_mask_fpclass_pd_mask (__mmask8 __U, __m512d __A, 2064 const int __imm) 2065 { 2066 return (__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) __A, 2067 __imm, __U); 2068 } 2069 2070 extern __inline __mmask8 2071 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2072 _mm512_fpclass_pd_mask (__m512d __A, const int __imm) 2073 { 2074 return (__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) __A, 2075 __imm, 2076 (__mmask8) -1); 2077 } 2078 2079 extern __inline __mmask16 2080 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2081 _mm512_mask_fpclass_ps_mask (__mmask16 __U, __m512 __A, 2082 const int __imm) 2083 { 2084 return (__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) __A, 2085 __imm, __U); 2086 } 2087 2088 extern __inline __mmask16 2089 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__)) 2090 _mm512_fpclass_ps_mask (__m512 __A, const int __imm) 2091 { 2092 return (__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) __A, 2093 __imm, 2094 (__mmask16) -1); 2095 } 2096 2097 #else 2098 #define _kshiftli_mask8(X, Y) \ 2099 ((__mmask8) __builtin_ia32_kshiftliqi ((__mmask8)(X), (__mmask8)(Y))) 2100 2101 #define _kshiftri_mask8(X, Y) \ 2102 ((__mmask8) __builtin_ia32_kshiftriqi ((__mmask8)(X), (__mmask8)(Y))) 2103 2104 #define _mm_range_sd(A, B, C) \ 2105 ((__m128d) __builtin_ia32_rangesd128_round ((__v2df)(__m128d)(A), \ 2106 (__v2df)(__m128d)(B), (int)(C), \ 2107 _MM_FROUND_CUR_DIRECTION)) 2108 2109 #define _mm_range_ss(A, B, C) \ 2110 ((__m128) __builtin_ia32_rangess128_round ((__v4sf)(__m128)(A), \ 2111 (__v4sf)(__m128)(B), (int)(C), \ 2112 _MM_FROUND_CUR_DIRECTION)) 2113 2114 #define _mm_range_round_sd(A, B, C, R) \ 2115 ((__m128d) __builtin_ia32_rangesd128_round ((__v2df)(__m128d)(A), \ 2116 (__v2df)(__m128d)(B), (int)(C), (R))) 2117 2118 #define _mm_range_round_ss(A, B, C, R) \ 2119 ((__m128) __builtin_ia32_rangess128_round ((__v4sf)(__m128)(A), \ 2120 (__v4sf)(__m128)(B), (int)(C), (R))) 2121 2122 #define _mm512_cvtt_roundpd_epi64(A, B) \ 2123 ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di) \ 2124 _mm512_setzero_si512 (), \ 2125 -1, (B))) 2126 2127 #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, B) \ 2128 ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di)(W), (U), (B))) 2129 2130 #define _mm512_maskz_cvtt_roundpd_epi64(U, A, B) \ 2131 ((__m512i)__builtin_ia32_cvttpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2132 2133 #define _mm512_cvtt_roundpd_epu64(A, B) \ 2134 ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2135 2136 #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, B) \ 2137 ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2138 2139 #define _mm512_maskz_cvtt_roundpd_epu64(U, A, B) \ 2140 ((__m512i)__builtin_ia32_cvttpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2141 2142 #define _mm512_cvtt_roundps_epi64(A, B) \ 2143 ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2144 2145 #define _mm512_mask_cvtt_roundps_epi64(W, U, A, B) \ 2146 ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)(W), (U), (B))) 2147 2148 #define _mm512_maskz_cvtt_roundps_epi64(U, A, B) \ 2149 ((__m512i)__builtin_ia32_cvttps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2150 2151 #define _mm512_cvtt_roundps_epu64(A, B) \ 2152 ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2153 2154 #define _mm512_mask_cvtt_roundps_epu64(W, U, A, B) \ 2155 ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2156 2157 #define _mm512_maskz_cvtt_roundps_epu64(U, A, B) \ 2158 ((__m512i)__builtin_ia32_cvttps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2159 2160 #define _mm512_cvt_roundpd_epi64(A, B) \ 2161 ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2162 2163 #define _mm512_mask_cvt_roundpd_epi64(W, U, A, B) \ 2164 ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)(W), (U), (B))) 2165 2166 #define _mm512_maskz_cvt_roundpd_epi64(U, A, B) \ 2167 ((__m512i)__builtin_ia32_cvtpd2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2168 2169 #define _mm512_cvt_roundpd_epu64(A, B) \ 2170 ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2171 2172 #define _mm512_mask_cvt_roundpd_epu64(W, U, A, B) \ 2173 ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2174 2175 #define _mm512_maskz_cvt_roundpd_epu64(U, A, B) \ 2176 ((__m512i)__builtin_ia32_cvtpd2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2177 2178 #define _mm512_cvt_roundps_epi64(A, B) \ 2179 ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2180 2181 #define _mm512_mask_cvt_roundps_epi64(W, U, A, B) \ 2182 ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)(W), (U), (B))) 2183 2184 #define _mm512_maskz_cvt_roundps_epi64(U, A, B) \ 2185 ((__m512i)__builtin_ia32_cvtps2qq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2186 2187 #define _mm512_cvt_roundps_epu64(A, B) \ 2188 ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), -1, (B))) 2189 2190 #define _mm512_mask_cvt_roundps_epu64(W, U, A, B) \ 2191 ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)(W), (U), (B))) 2192 2193 #define _mm512_maskz_cvt_roundps_epu64(U, A, B) \ 2194 ((__m512i)__builtin_ia32_cvtps2uqq512_mask ((A), (__v8di)_mm512_setzero_si512 (), (U), (B))) 2195 2196 #define _mm512_cvt_roundepi64_ps(A, B) \ 2197 ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), -1, (B))) 2198 2199 #define _mm512_mask_cvt_roundepi64_ps(W, U, A, B) \ 2200 ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (W), (U), (B))) 2201 2202 #define _mm512_maskz_cvt_roundepi64_ps(U, A, B) \ 2203 ((__m256)__builtin_ia32_cvtqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), (U), (B))) 2204 2205 #define _mm512_cvt_roundepu64_ps(A, B) \ 2206 ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), -1, (B))) 2207 2208 #define _mm512_mask_cvt_roundepu64_ps(W, U, A, B) \ 2209 ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (W), (U), (B))) 2210 2211 #define _mm512_maskz_cvt_roundepu64_ps(U, A, B) \ 2212 ((__m256)__builtin_ia32_cvtuqq2ps512_mask ((__v8di)(A), (__v8sf)_mm256_setzero_ps (), (U), (B))) 2213 2214 #define _mm512_cvt_roundepi64_pd(A, B) \ 2215 ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), -1, (B))) 2216 2217 #define _mm512_mask_cvt_roundepi64_pd(W, U, A, B) \ 2218 ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (W), (U), (B))) 2219 2220 #define _mm512_maskz_cvt_roundepi64_pd(U, A, B) \ 2221 ((__m512d)__builtin_ia32_cvtqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), (U), (B))) 2222 2223 #define _mm512_cvt_roundepu64_pd(A, B) \ 2224 ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), -1, (B))) 2225 2226 #define _mm512_mask_cvt_roundepu64_pd(W, U, A, B) \ 2227 ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (W), (U), (B))) 2228 2229 #define _mm512_maskz_cvt_roundepu64_pd(U, A, B) \ 2230 ((__m512d)__builtin_ia32_cvtuqq2pd512_mask ((__v8di)(A), (__v8df)_mm512_setzero_pd (), (U), (B))) 2231 2232 #define _mm512_reduce_pd(A, B) \ 2233 ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ 2234 (int)(B), (__v8df)_mm512_setzero_pd (), (__mmask8)-1)) 2235 2236 #define _mm512_mask_reduce_pd(W, U, A, B) \ 2237 ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ 2238 (int)(B), (__v8df)(__m512d)(W), (__mmask8)(U))) 2239 2240 #define _mm512_maskz_reduce_pd(U, A, B) \ 2241 ((__m512d) __builtin_ia32_reducepd512_mask ((__v8df)(__m512d)(A), \ 2242 (int)(B), (__v8df)_mm512_setzero_pd (), (__mmask8)(U))) 2243 2244 #define _mm512_reduce_ps(A, B) \ 2245 ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ 2246 (int)(B), (__v16sf)_mm512_setzero_ps (), (__mmask16)-1)) 2247 2248 #define _mm512_mask_reduce_ps(W, U, A, B) \ 2249 ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ 2250 (int)(B), (__v16sf)(__m512)(W), (__mmask16)(U))) 2251 2252 #define _mm512_maskz_reduce_ps(U, A, B) \ 2253 ((__m512) __builtin_ia32_reduceps512_mask ((__v16sf)(__m512)(A), \ 2254 (int)(B), (__v16sf)_mm512_setzero_ps (), (__mmask16)(U))) 2255 2256 #define _mm512_extractf32x8_ps(X, C) \ 2257 ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ 2258 (int) (C), (__v8sf)(__m256) _mm256_setzero_ps (), (__mmask8)-1)) 2259 2260 #define _mm512_mask_extractf32x8_ps(W, U, X, C) \ 2261 ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ 2262 (int) (C), (__v8sf)(__m256) (W), (__mmask8) (U))) 2263 2264 #define _mm512_maskz_extractf32x8_ps(U, X, C) \ 2265 ((__m256) __builtin_ia32_extractf32x8_mask ((__v16sf)(__m512) (X), \ 2266 (int) (C), (__v8sf)(__m256) _mm256_setzero_ps (), (__mmask8) (U))) 2267 2268 #define _mm512_extractf64x2_pd(X, C) \ 2269 ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ 2270 (int) (C), (__v2df)(__m128d) _mm_setzero_pd (), (__mmask8)-1)) 2271 2272 #define _mm512_mask_extractf64x2_pd(W, U, X, C) \ 2273 ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ 2274 (int) (C), (__v2df)(__m128d) (W), (__mmask8) (U))) 2275 2276 #define _mm512_maskz_extractf64x2_pd(U, X, C) \ 2277 ((__m128d) __builtin_ia32_extractf64x2_512_mask ((__v8df)(__m512d) (X),\ 2278 (int) (C), (__v2df)(__m128d) _mm_setzero_pd (), (__mmask8) (U))) 2279 2280 #define _mm512_extracti32x8_epi32(X, C) \ 2281 ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ 2282 (int) (C), (__v8si)(__m256i) _mm256_setzero_si256 (), (__mmask8)-1)) 2283 2284 #define _mm512_mask_extracti32x8_epi32(W, U, X, C) \ 2285 ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ 2286 (int) (C), (__v8si)(__m256i) (W), (__mmask8) (U))) 2287 2288 #define _mm512_maskz_extracti32x8_epi32(U, X, C) \ 2289 ((__m256i) __builtin_ia32_extracti32x8_mask ((__v16si)(__m512i) (X), \ 2290 (int) (C), (__v8si)(__m256i) _mm256_setzero_si256 (), (__mmask8) (U))) 2291 2292 #define _mm512_extracti64x2_epi64(X, C) \ 2293 ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ 2294 (int) (C), (__v2di)(__m128i) _mm_setzero_si128 (), (__mmask8)-1)) 2295 2296 #define _mm512_mask_extracti64x2_epi64(W, U, X, C) \ 2297 ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ 2298 (int) (C), (__v2di)(__m128i) (W), (__mmask8) (U))) 2299 2300 #define _mm512_maskz_extracti64x2_epi64(U, X, C) \ 2301 ((__m128i) __builtin_ia32_extracti64x2_512_mask ((__v8di)(__m512i) (X),\ 2302 (int) (C), (__v2di)(__m128i) _mm_setzero_si128 (), (__mmask8) (U))) 2303 2304 #define _mm512_range_pd(A, B, C) \ 2305 ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2306 (__v8df)(__m512d)(B), (int)(C), \ 2307 (__v8df)_mm512_setzero_pd (), (__mmask8)-1, _MM_FROUND_CUR_DIRECTION)) 2308 2309 #define _mm512_mask_range_pd(W, U, A, B, C) \ 2310 ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2311 (__v8df)(__m512d)(B), (int)(C), \ 2312 (__v8df)(__m512d)(W), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2313 2314 #define _mm512_maskz_range_pd(U, A, B, C) \ 2315 ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2316 (__v8df)(__m512d)(B), (int)(C), \ 2317 (__v8df)_mm512_setzero_pd (), (__mmask8)(U), _MM_FROUND_CUR_DIRECTION)) 2318 2319 #define _mm512_range_ps(A, B, C) \ 2320 ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2321 (__v16sf)(__m512)(B), (int)(C), \ 2322 (__v16sf)_mm512_setzero_ps (), (__mmask16)-1, _MM_FROUND_CUR_DIRECTION)) 2323 2324 #define _mm512_mask_range_ps(W, U, A, B, C) \ 2325 ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2326 (__v16sf)(__m512)(B), (int)(C), \ 2327 (__v16sf)(__m512)(W), (__mmask16)(U), _MM_FROUND_CUR_DIRECTION)) 2328 2329 #define _mm512_maskz_range_ps(U, A, B, C) \ 2330 ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2331 (__v16sf)(__m512)(B), (int)(C), \ 2332 (__v16sf)_mm512_setzero_ps (), (__mmask16)(U), _MM_FROUND_CUR_DIRECTION)) 2333 2334 #define _mm512_range_round_pd(A, B, C, R) \ 2335 ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2336 (__v8df)(__m512d)(B), (int)(C), \ 2337 (__v8df)_mm512_setzero_pd (), (__mmask8)-1, (R))) 2338 2339 #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ 2340 ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2341 (__v8df)(__m512d)(B), (int)(C), \ 2342 (__v8df)(__m512d)(W), (__mmask8)(U), (R))) 2343 2344 #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ 2345 ((__m512d) __builtin_ia32_rangepd512_mask ((__v8df)(__m512d)(A), \ 2346 (__v8df)(__m512d)(B), (int)(C), \ 2347 (__v8df)_mm512_setzero_pd (), (__mmask8)(U), (R))) 2348 2349 #define _mm512_range_round_ps(A, B, C, R) \ 2350 ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2351 (__v16sf)(__m512)(B), (int)(C), \ 2352 (__v16sf)_mm512_setzero_ps (), (__mmask16)-1, (R))) 2353 2354 #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ 2355 ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2356 (__v16sf)(__m512)(B), (int)(C), \ 2357 (__v16sf)(__m512)(W), (__mmask16)(U), (R))) 2358 2359 #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ 2360 ((__m512) __builtin_ia32_rangeps512_mask ((__v16sf)(__m512)(A), \ 2361 (__v16sf)(__m512)(B), (int)(C), \ 2362 (__v16sf)_mm512_setzero_ps (), (__mmask16)(U), (R))) 2363 2364 #define _mm512_insertf64x2(X, Y, C) \ 2365 ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ 2366 (__v2df)(__m128d) (Y), (int) (C), (__v8df)(__m512d) (X), \ 2367 (__mmask8)-1)) 2368 2369 #define _mm512_mask_insertf64x2(W, U, X, Y, C) \ 2370 ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ 2371 (__v2df)(__m128d) (Y), (int) (C), (__v8df)(__m512d) (W), \ 2372 (__mmask8) (U))) 2373 2374 #define _mm512_maskz_insertf64x2(U, X, Y, C) \ 2375 ((__m512d) __builtin_ia32_insertf64x2_512_mask ((__v8df)(__m512d) (X),\ 2376 (__v2df)(__m128d) (Y), (int) (C), \ 2377 (__v8df)(__m512d) _mm512_setzero_pd (), (__mmask8) (U))) 2378 2379 #define _mm512_inserti64x2(X, Y, C) \ 2380 ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ 2381 (__v2di)(__m128i) (Y), (int) (C), (__v8di)(__m512i) (X), (__mmask8)-1)) 2382 2383 #define _mm512_mask_inserti64x2(W, U, X, Y, C) \ 2384 ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ 2385 (__v2di)(__m128i) (Y), (int) (C), (__v8di)(__m512i) (W), \ 2386 (__mmask8) (U))) 2387 2388 #define _mm512_maskz_inserti64x2(U, X, Y, C) \ 2389 ((__m512i) __builtin_ia32_inserti64x2_512_mask ((__v8di)(__m512i) (X),\ 2390 (__v2di)(__m128i) (Y), (int) (C), \ 2391 (__v8di)(__m512i) _mm512_setzero_si512 (), (__mmask8) (U))) 2392 2393 #define _mm512_insertf32x8(X, Y, C) \ 2394 ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ 2395 (__v8sf)(__m256) (Y), (int) (C),\ 2396 (__v16sf)(__m512)_mm512_setzero_ps (),\ 2397 (__mmask16)-1)) 2398 2399 #define _mm512_mask_insertf32x8(W, U, X, Y, C) \ 2400 ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ 2401 (__v8sf)(__m256) (Y), (int) (C),\ 2402 (__v16sf)(__m512)(W),\ 2403 (__mmask16)(U))) 2404 2405 #define _mm512_maskz_insertf32x8(U, X, Y, C) \ 2406 ((__m512) __builtin_ia32_insertf32x8_mask ((__v16sf)(__m512) (X), \ 2407 (__v8sf)(__m256) (Y), (int) (C),\ 2408 (__v16sf)(__m512)_mm512_setzero_ps (),\ 2409 (__mmask16)(U))) 2410 2411 #define _mm512_inserti32x8(X, Y, C) \ 2412 ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ 2413 (__v8si)(__m256i) (Y), (int) (C),\ 2414 (__v16si)(__m512i)_mm512_setzero_si512 (),\ 2415 (__mmask16)-1)) 2416 2417 #define _mm512_mask_inserti32x8(W, U, X, Y, C) \ 2418 ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ 2419 (__v8si)(__m256i) (Y), (int) (C),\ 2420 (__v16si)(__m512i)(W),\ 2421 (__mmask16)(U))) 2422 2423 #define _mm512_maskz_inserti32x8(U, X, Y, C) \ 2424 ((__m512i) __builtin_ia32_inserti32x8_mask ((__v16si)(__m512i) (X), \ 2425 (__v8si)(__m256i) (Y), (int) (C),\ 2426 (__v16si)(__m512i)_mm512_setzero_si512 (),\ 2427 (__mmask16)(U))) 2428 2429 #define _mm_fpclass_ss_mask(X, C) \ 2430 ((__mmask8) __builtin_ia32_fpclassss ((__v4sf) (__m128) (X), (int) (C))) \ 2431 2432 #define _mm_fpclass_sd_mask(X, C) \ 2433 ((__mmask8) __builtin_ia32_fpclasssd ((__v2df) (__m128d) (X), (int) (C))) \ 2434 2435 #define _mm512_mask_fpclass_pd_mask(u, X, C) \ 2436 ((__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) (__m512d) (X), \ 2437 (int) (C), (__mmask8)(u))) 2438 2439 #define _mm512_mask_fpclass_ps_mask(u, x, c) \ 2440 ((__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) (__m512) (x),\ 2441 (int) (c),(__mmask8)(u))) 2442 2443 #define _mm512_fpclass_pd_mask(X, C) \ 2444 ((__mmask8) __builtin_ia32_fpclasspd512_mask ((__v8df) (__m512d) (X), \ 2445 (int) (C), (__mmask8)-1)) 2446 2447 #define _mm512_fpclass_ps_mask(x, c) \ 2448 ((__mmask16) __builtin_ia32_fpclassps512_mask ((__v16sf) (__m512) (x),\ 2449 (int) (c),(__mmask8)-1)) 2450 2451 #define _mm_reduce_sd(A, B, C) \ 2452 ((__m128d) __builtin_ia32_reducesd ((__v2df)(__m128d)(A), \ 2453 (__v2df)(__m128d)(B), (int)(C))) \ 2454 2455 #define _mm_reduce_ss(A, B, C) \ 2456 ((__m128) __builtin_ia32_reducess ((__v4sf)(__m128)(A), \ 2457 (__v4sf)(__m128)(A), (int)(C))) \ 2458 2459 #endif 2460 2461 #ifdef __DISABLE_AVX512DQ__ 2462 #undef __DISABLE_AVX512DQ__ 2463 #pragma GCC pop_options 2464 #endif /* __DISABLE_AVX512DQ__ */ 2465 2466 #endif /* _AVX512DQINTRIN_H_INCLUDED */ 2467
Indexes created Sat Apr 25 00:22:52 UTC 2026