gcc/config/i386/avx10_2-512bf16intrin.h

   1 /* Copyright (C) 2024-2025 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 <avx10_2-512bf16intrin.h> directly; include <immintrin.h> instead."
  26 #endif
  27
  28 #ifndef _AVX10_2_512BF16INTRIN_H_INCLUDED
  29 #define _AVX10_2_512BF16INTRIN_H_INCLUDED
  30
  31 #if !defined (__AVX10_2_512__)
  32 #pragma GCC push_options
  33 #pragma GCC target("avx10.2-512")
  34 #define __DISABLE_AVX10_2_512__
  35 #endif /* __AVX10_2_512__ */
  36
  37 extern __inline__ __m512bh
  38 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  39 _mm512_add_pbh (__m512bh __A, __m512bh __B)
  40 {
  41   return (__m512bh) __builtin_ia32_addbf16512 (__A, __B);
  42 }
  43
  44 extern __inline__ __m512bh
  45 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  46 _mm512_mask_add_pbh (__m512bh __W, __mmask32 __U,
  47                      __m512bh __A, __m512bh __B)
  48 {
  49   return (__m512bh)
  50     __builtin_ia32_addbf16512_mask (__A, __B, __W, __U);
  51 }
  52
  53 extern __inline__ __m512bh
  54 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  55 _mm512_maskz_add_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
  56 {
  57   return (__m512bh)
  58     __builtin_ia32_addbf16512_mask (__A, __B,
  59                                     (__v32bf) _mm512_setzero_si512 (),
  60                                     __U);
  61 }
  62
  63 extern __inline__ __m512bh
  64 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  65 _mm512_sub_pbh (__m512bh __A, __m512bh __B)
  66 {
  67   return (__m512bh) __builtin_ia32_subbf16512 (__A, __B);
  68 }
  69
  70 extern __inline__ __m512bh
  71 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  72 _mm512_mask_sub_pbh (__m512bh __W, __mmask32 __U,
  73                      __m512bh __A, __m512bh __B)
  74 {
  75   return (__m512bh)
  76     __builtin_ia32_subbf16512_mask (__A, __B, __W, __U);
  77 }
  78
  79 extern __inline__ __m512bh
  80 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  81 _mm512_maskz_sub_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
  82 {
  83   return (__m512bh)
  84     __builtin_ia32_subbf16512_mask (__A, __B,
  85                                     (__v32bf) _mm512_setzero_si512 (),
  86                                     __U);
  87 }
  88
  89 extern __inline__ __m512bh
  90 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  91 _mm512_mul_pbh (__m512bh __A, __m512bh __B)
  92 {
  93   return (__m512bh) __builtin_ia32_mulbf16512 (__A, __B);
  94 }
  95
  96 extern __inline__ __m512bh
  97 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
  98 _mm512_mask_mul_pbh (__m512bh __W, __mmask32 __U,
  99                      __m512bh __A, __m512bh __B)
 100 {
 101   return (__m512bh)
 102     __builtin_ia32_mulbf16512_mask (__A, __B, __W, __U);
 103 }
 104
 105 extern __inline__ __m512bh
 106 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 107 _mm512_maskz_mul_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
 108 {
 109   return (__m512bh)
 110     __builtin_ia32_mulbf16512_mask (__A, __B,
 111                                     (__v32bf) _mm512_setzero_si512 (),
 112                                     __U);
 113 }
 114
 115 extern __inline__ __m512bh
 116 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 117 _mm512_div_pbh (__m512bh __A, __m512bh __B)
 118 {
 119   return (__m512bh) __builtin_ia32_divbf16512 (__A, __B);
 120 }
 121
 122 extern __inline__ __m512bh
 123 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 124 _mm512_mask_div_pbh (__m512bh __W, __mmask32 __U,
 125                      __m512bh __A, __m512bh __B)
 126 {
 127   return (__m512bh)
 128     __builtin_ia32_divbf16512_mask (__A, __B, __W, __U);
 129 }
 130
 131 extern __inline__ __m512bh
 132 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 133 _mm512_maskz_div_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
 134 {
 135   return (__m512bh)
 136     __builtin_ia32_divbf16512_mask (__A, __B,
 137                                     (__v32bf) _mm512_setzero_si512 (),
 138                                     __U);
 139 }
 140
 141 extern __inline__ __m512bh
 142 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 143 _mm512_max_pbh (__m512bh __A, __m512bh __B)
 144 {
 145   return (__m512bh) __builtin_ia32_maxbf16512 (__A, __B);
 146 }
 147
 148 extern __inline__ __m512bh
 149 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 150 _mm512_mask_max_pbh (__m512bh __W, __mmask32 __U,
 151                      __m512bh __A, __m512bh __B)
 152 {
 153   return (__m512bh)
 154     __builtin_ia32_maxbf16512_mask (__A, __B, __W, __U);
 155 }
 156
 157 extern __inline__ __m512bh
 158 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 159 _mm512_maskz_max_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
 160 {
 161   return (__m512bh)
 162     __builtin_ia32_maxbf16512_mask (__A, __B,
 163                                     (__v32bf) _mm512_setzero_si512 (),
 164                                     __U);
 165 }
 166
 167 extern __inline__ __m512bh
 168 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 169 _mm512_min_pbh (__m512bh __A, __m512bh __B)
 170 {
 171   return (__m512bh) __builtin_ia32_minbf16512 (__A, __B);
 172 }
 173
 174 extern __inline__ __m512bh
 175 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 176 _mm512_mask_min_pbh (__m512bh __W, __mmask32 __U,
 177                      __m512bh __A, __m512bh __B)
 178 {
 179   return (__m512bh)
 180     __builtin_ia32_minbf16512_mask (__A, __B, __W, __U);
 181 }
 182
 183 extern __inline__ __m512bh
 184 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 185 _mm512_maskz_min_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
 186 {
 187   return (__m512bh)
 188     __builtin_ia32_minbf16512_mask (__A, __B,
 189                                     (__v32bf) _mm512_setzero_si512 (),
 190                                     __U);
 191 }
 192
 193 extern __inline__ __m512bh
 194 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 195 _mm512_scalef_pbh (__m512bh __A, __m512bh __B)
 196 {
 197   return (__m512bh) __builtin_ia32_scalefbf16512 (__A, __B);
 198 }
 199
 200 extern __inline__ __m512bh
 201 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 202 _mm512_mask_scalef_pbh (__m512bh __W, __mmask32 __U,
 203                         __m512bh __A, __m512bh __B)
 204 {
 205   return (__m512bh)
 206     __builtin_ia32_scalefbf16512_mask (__A, __B, __W, __U);
 207 }
 208
 209 extern __inline__ __m512bh
 210 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 211 _mm512_maskz_scalef_pbh (__mmask32 __U, __m512bh __A, __m512bh __B)
 212 {
 213   return (__m512bh)
 214     __builtin_ia32_scalefbf16512_mask (__A, __B,
 215                                        (__v32bf) _mm512_setzero_si512 (),
 216                                        __U);
 217 }
 218
 219 extern __inline__ __m512bh
 220 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 221 _mm512_fmadd_pbh (__m512bh __A, __m512bh __B, __m512bh __C)
 222 {
 223   return (__m512bh)
 224     __builtin_ia32_fmaddbf16512_mask (__A, __B, __C, (__mmask32) -1);
 225 }
 226
 227 extern __inline__ __m512bh
 228 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 229 _mm512_mask_fmadd_pbh (__m512bh __A, __mmask32 __U,
 230                        __m512bh __B, __m512bh __C)
 231 {
 232   return (__m512bh)
 233     __builtin_ia32_fmaddbf16512_mask (__A, __B, __C, __U);
 234 }
 235
 236 extern __inline__ __m512bh
 237 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 238 _mm512_mask3_fmadd_pbh (__m512bh __A, __m512bh __B,
 239                         __m512bh __C, __mmask32 __U)
 240 {
 241   return (__m512bh)
 242     __builtin_ia32_fmaddbf16512_mask3 (__A, __B, __C, __U);
 243 }
 244
 245 extern __inline__ __m512bh
 246 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 247 _mm512_maskz_fmadd_pbh (__mmask32 __U, __m512bh __A,
 248                           __m512bh __B, __m512bh __C)
 249 {
 250   return (__m512bh)
 251     __builtin_ia32_fmaddbf16512_maskz (__A, __B, __C, __U);
 252 }
 253
 254 extern __inline__ __m512bh
 255 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 256 _mm512_fmsub_pbh (__m512bh __A, __m512bh __B, __m512bh __C)
 257 {
 258   return (__m512bh)
 259     __builtin_ia32_fmsubbf16512_mask (__A, __B, __C, (__mmask32) -1);
 260 }
 261
 262 extern __inline__ __m512bh
 263 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 264 _mm512_mask_fmsub_pbh (__m512bh __A, __mmask32 __U,
 265                        __m512bh __B, __m512bh __C)
 266 {
 267   return (__m512bh)
 268     __builtin_ia32_fmsubbf16512_mask (__A, __B, __C, __U);
 269 }
 270
 271 extern __inline__ __m512bh
 272 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 273 _mm512_mask3_fmsub_pbh (__m512bh __A, __m512bh __B,
 274                         __m512bh __C, __mmask32 __U)
 275 {
 276   return (__m512bh)
 277     __builtin_ia32_fmsubbf16512_mask3 (__A, __B, __C, __U);
 278 }
 279
 280 extern __inline__ __m512bh
 281 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 282 _mm512_maskz_fmsub_pbh (__mmask32 __U, __m512bh __A,
 283                         __m512bh __B, __m512bh __C)
 284 {
 285   return (__m512bh)
 286     __builtin_ia32_fmsubbf16512_maskz (__A, __B, __C, __U);
 287 }
 288
 289 extern __inline__ __m512bh
 290 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 291 _mm512_fnmadd_pbh (__m512bh __A, __m512bh __B, __m512bh __C)
 292 {
 293   return (__m512bh)
 294     __builtin_ia32_fnmaddbf16512_mask (__A, __B, __C, (__mmask32) -1);
 295 }
 296
 297 extern __inline__ __m512bh
 298 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 299 _mm512_mask_fnmadd_pbh (__m512bh __A, __mmask32 __U,
 300                         __m512bh __B, __m512bh __C)
 301 {
 302   return (__m512bh)
 303     __builtin_ia32_fnmaddbf16512_mask (__A, __B, __C, __U);
 304 }
 305
 306 extern __inline__ __m512bh
 307 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 308 _mm512_mask3_fnmadd_pbh (__m512bh __A, __m512bh __B,
 309                          __m512bh __C, __mmask32 __U)
 310 {
 311   return (__m512bh)
 312     __builtin_ia32_fnmaddbf16512_mask3 (__A, __B, __C, __U);
 313 }
 314
 315 extern __inline__ __m512bh
 316 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 317 _mm512_maskz_fnmadd_pbh (__mmask32 __U, __m512bh __A,
 318                          __m512bh __B, __m512bh __C)
 319 {
 320   return (__m512bh)
 321     __builtin_ia32_fnmaddbf16512_maskz (__A, __B, __C, __U);
 322 }
 323
 324 extern __inline__ __m512bh
 325 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 326 _mm512_fnmsub_pbh (__m512bh __A, __m512bh __B, __m512bh __C)
 327 {
 328   return (__m512bh)
 329     __builtin_ia32_fnmsubbf16512_mask (__A, __B, __C, (__mmask32) -1);
 330 }
 331
 332 extern __inline__ __m512bh
 333 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 334 _mm512_mask_fnmsub_pbh (__m512bh __A, __mmask32 __U,
 335                         __m512bh __B, __m512bh __C)
 336 {
 337   return (__m512bh)
 338     __builtin_ia32_fnmsubbf16512_mask (__A, __B, __C, __U);
 339 }
 340
 341 extern __inline__ __m512bh
 342 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 343 _mm512_mask3_fnmsub_pbh (__m512bh __A, __m512bh __B,
 344                          __m512bh __C, __mmask32 __U)
 345 {
 346   return (__m512bh)
 347     __builtin_ia32_fnmsubbf16512_mask3 (__A, __B, __C, __U);
 348 }
 349
 350 extern __inline__ __m512bh
 351 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 352 _mm512_maskz_fnmsub_pbh (__mmask32 __U, __m512bh __A,
 353                          __m512bh __B, __m512bh __C)
 354 {
 355   return (__m512bh)
 356     __builtin_ia32_fnmsubbf16512_maskz (__A, __B, __C, __U);
 357 }
 358
 359 extern __inline__ __m512bh
 360 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 361 _mm512_rsqrt_pbh (__m512bh __A)
 362 {
 363   return (__m512bh)
 364     __builtin_ia32_rsqrtbf16512_mask (__A,
 365                                       (__v32bf) _mm512_setzero_si512 (),
 366                                       (__mmask32) -1);
 367
 368 }
 369
 370 extern __inline__ __m512bh
 371 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 372 _mm512_mask_rsqrt_pbh (__m512bh __W, __mmask32 __U, __m512bh __A)
 373 {
 374   return (__m512bh)
 375     __builtin_ia32_rsqrtbf16512_mask (__A,  __W,  __U);
 376 }
 377
 378 extern __inline__ __m512bh
 379 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 380 _mm512_maskz_rsqrt_pbh (__mmask32 __U, __m512bh __A)
 381 {
 382   return (__m512bh)
 383     __builtin_ia32_rsqrtbf16512_mask (__A,
 384                                       (__v32bf) _mm512_setzero_si512 (),
 385                                       __U);
 386 }
 387
 388 extern __inline__ __m512bh
 389 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 390 _mm512_sqrt_pbh (__m512bh __A)
 391 {
 392   return (__m512bh)
 393     __builtin_ia32_sqrtbf16512_mask (__A,
 394                                      (__v32bf) _mm512_setzero_si512 (),
 395                                      (__mmask32) -1);
 396 }
 397
 398 extern __inline__ __m512bh
 399 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 400 _mm512_mask_sqrt_pbh (__m512bh __W, __mmask32 __U, __m512bh __A)
 401 {
 402   return (__m512bh)
 403     __builtin_ia32_sqrtbf16512_mask (__A,  __W,  __U);
 404 }
 405
 406 extern __inline__ __m512bh
 407 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 408 _mm512_maskz_sqrt_pbh (__mmask32 __U, __m512bh __A)
 409 {
 410   return (__m512bh)
 411     __builtin_ia32_sqrtbf16512_mask (__A,
 412                                      (__v32bf) _mm512_setzero_si512 (),
 413                                      __U);
 414 }
 415
 416 extern __inline__ __m512bh
 417 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 418 _mm512_rcp_pbh (__m512bh __A)
 419 {
 420   return (__m512bh)
 421     __builtin_ia32_rcpbf16512_mask (__A,
 422                                     (__v32bf) _mm512_setzero_si512 (),
 423                                     (__mmask32) -1);
 424 }
 425
 426 extern __inline__ __m512bh
 427 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 428 _mm512_mask_rcp_pbh (__m512bh __W, __mmask32 __U, __m512bh __A)
 429 {
 430   return (__m512bh)
 431     __builtin_ia32_rcpbf16512_mask (__A,  __W,  __U);
 432 }
 433
 434 extern __inline__ __m512bh
 435 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 436 _mm512_maskz_rcp_pbh (__mmask32 __U, __m512bh __A)
 437 {
 438   return (__m512bh)
 439     __builtin_ia32_rcpbf16512_mask (__A,
 440                                     (__v32bf) _mm512_setzero_si512 (),
 441                                     __U);
 442 }
 443
 444 extern __inline__ __m512bh
 445 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 446 _mm512_getexp_pbh (__m512bh __A)
 447 {
 448   return (__m512bh)
 449     __builtin_ia32_getexpbf16512_mask (__A,
 450                                        (__v32bf) _mm512_setzero_si512 (),
 451                                        (__mmask32) -1);
 452 }
 453
 454 extern __inline__ __m512bh
 455 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 456 _mm512_mask_getexp_pbh (__m512bh __W, __mmask32 __U, __m512bh __A)
 457 {
 458   return (__m512bh) __builtin_ia32_getexpbf16512_mask (__A,  __W,  __U);
 459 }
 460
 461 extern __inline__ __m512bh
 462 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 463 _mm512_maskz_getexp_pbh (__mmask32 __U, __m512bh __A)
 464 {
 465   return (__m512bh)
 466     __builtin_ia32_getexpbf16512_mask (__A,
 467                                        (__v32bf) _mm512_setzero_si512 (),
 468                                        __U);
 469 }
 470
 471 /* Intrinsics vrndscalebf16.  */
 472 #ifdef __OPTIMIZE__
 473 extern __inline__ __m512bh
 474 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 475 _mm512_roundscale_pbh (__m512bh __A, int B)
 476 {
 477   return (__m512bh)
 478     __builtin_ia32_rndscalebf16512_mask (__A, B,
 479                                          (__v32bf) _mm512_setzero_si512 (),
 480                                          (__mmask32) -1);
 481 }
 482
 483 extern __inline__ __m512bh
 484 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 485 _mm512_mask_roundscale_pbh (__m512bh __W, __mmask32 __U, __m512bh __A, int B)
 486 {
 487   return (__m512bh)
 488     __builtin_ia32_rndscalebf16512_mask (__A, B, __W,  __U);
 489 }
 490
 491 extern __inline__ __m512bh
 492 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 493 _mm512_maskz_roundscale_pbh (__mmask32 __U, __m512bh __A, int B)
 494 {
 495   return (__m512bh)
 496     __builtin_ia32_rndscalebf16512_mask (__A, B,
 497                                          (__v32bf) _mm512_setzero_si512 (),
 498                                          __U);
 499 }
 500
 501 #else
 502 #define _mm512_roundscale_pbh(A, B)                                           \
 503   (__builtin_ia32_rndscalebf16512_mask ((A), (B),                             \
 504                                         (__v32bf) _mm512_setzero_si512 (),    \
 505                                         (__mmask32) -1))
 506
 507 #define _mm512_mask_roundscale_pbh(A, B, C, D)                        \
 508   (__builtin_ia32_rndscalebf16512_mask ((C), (D), (A), (B)))
 509
 510 #define _mm512_maskz_roundscale_pbh(A, B, C)                                  \
 511   (__builtin_ia32_rndscalebf16512_mask ((B), (C),                             \
 512                                         (__v32bf) _mm512_setzero_si512 (),    \
 513                                         (A)))
 514
 515 #endif /* __OPTIMIZE__ */
 516
 517 /* Intrinsics vreducebf16.  */
 518 #ifdef __OPTIMIZE__
 519 extern __inline__ __m512bh
 520 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 521 _mm512_reduce_pbh (__m512bh __A, int B)
 522 {
 523   return (__m512bh)
 524     __builtin_ia32_reducebf16512_mask (__A, B,
 525                                        (__v32bf) _mm512_setzero_si512 (),
 526                                        (__mmask32) -1);
 527 }
 528
 529 extern __inline__ __m512bh
 530 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 531 _mm512_mask_reduce_pbh (__m512bh __W, __mmask32 __U,
 532                         __m512bh __A, int B)
 533 {
 534   return (__m512bh)
 535     __builtin_ia32_reducebf16512_mask (__A, B, __W,  __U);
 536 }
 537
 538 extern __inline__ __m512bh
 539 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 540 _mm512_maskz_reduce_pbh (__mmask32 __U, __m512bh __A, int B)
 541 {
 542   return (__m512bh)
 543     __builtin_ia32_reducebf16512_mask (__A, B,
 544                                           (__v32bf) _mm512_setzero_si512 (),
 545                                           __U);
 546 }
 547
 548 #else
 549 #define _mm512_reduce_pbh(A, B)                                       \
 550   (__builtin_ia32_reducebf16512_mask ((A), (B),                       \
 551                                       (__v32bf) _mm512_setzero_si512 (),   \
 552                                       (__mmask32) -1))
 553
 554 #define _mm512_mask_reduce_pbh(A, B, C, D)                                    \
 555   (__builtin_ia32_reducebf16512_mask ((C), (D), (A), (B)))
 556
 557 #define _mm512_maskz_reduce_pbh(A, B, C)                                      \
 558   (__builtin_ia32_reducebf16512_mask ((B), (C),                       \
 559                                       (__v32bf) _mm512_setzero_si512 (),      \
 560                                       (A)))
 561
 562 #endif /* __OPTIMIZE__ */
 563
 564 /* Intrinsics vgetmantbf16.  */
 565 #ifdef __OPTIMIZE__
 566 extern __inline__ __m512bh
 567 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 568 _mm512_getmant_pbh (__m512bh __A, _MM_MANTISSA_NORM_ENUM __B,
 569                     _MM_MANTISSA_SIGN_ENUM __C)
 570 {
 571   return (__m512bh)
 572     __builtin_ia32_getmantbf16512_mask (__A, (int) (__C << 2) | __B,
 573                                         (__v32bf) _mm512_setzero_si512 (),
 574                                         (__mmask32) -1);
 575 }
 576
 577 extern __inline__ __m512bh
 578 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 579 _mm512_mask_getmant_pbh (__m512bh __W, __mmask32 __U, __m512bh __A,
 580                          _MM_MANTISSA_NORM_ENUM __B,
 581                          _MM_MANTISSA_SIGN_ENUM __C)
 582 {
 583   return (__m512bh)
 584     __builtin_ia32_getmantbf16512_mask (__A, (int) (__C << 2) | __B,
 585                                         __W, __U);
 586 }
 587
 588 extern __inline__ __m512bh
 589 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 590 _mm512_maskz_getmant_pbh (__mmask32 __U, __m512bh __A,
 591                           _MM_MANTISSA_NORM_ENUM __B,
 592                           _MM_MANTISSA_SIGN_ENUM __C)
 593 {
 594   return (__m512bh)
 595     __builtin_ia32_getmantbf16512_mask (__A, (int) (__C << 2) | __B,
 596                                         (__v32bf) _mm512_setzero_si512 (),
 597                                         __U);
 598 }
 599
 600 #else
 601 #define _mm512_getmant_pbh(A, B, C)                                           \
 602   (__builtin_ia32_getmantbf16512_mask ((A), (int)(((C)<<2) | (B)),            \
 603                                        (__v32bf) _mm512_setzero_si512 (),     \
 604                                        (__mmask32) -1))
 605
 606 #define _mm512_mask_getmant_pbh(A, B, C, D, E)                                \
 607   (__builtin_ia32_getmantbf16512_mask ((C), (int)(((D)<<2) | (E)), (A), (B)))
 608
 609 #define _mm512_maskz_getmant_pbh(A, B, C, D)                                  \
 610   (__builtin_ia32_getmantbf16512_mask ((B), (int)(((C)<<2) | (D)),            \
 611                                        (__v32bf) _mm512_setzero_si512 (),     \
 612                                           (A)))
 613
 614 #endif /* __OPTIMIZE__ */
 615
 616 /* Intrinsics vfpclassbf16.  */
 617 #ifdef __OPTIMIZE__
 618 extern __inline __mmask32
 619 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 620 _mm512_mask_fpclass_pbh_mask (__mmask32 __U, __m512bh __A,
 621                               const int __imm)
 622 {
 623   return (__mmask32)
 624     __builtin_ia32_fpclassbf16512_mask (__A, __imm, __U);
 625 }
 626
 627 extern __inline __mmask32
 628 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 629 _mm512_fpclass_pbh_mask (__m512bh __A, const int __imm)
 630 {
 631   return (__mmask32)
 632     __builtin_ia32_fpclassbf16512_mask (__A, __imm,
 633                                         (__mmask32) -1);
 634 }
 635
 636 #else
 637 #define _mm512_mask_fpclass_pbh_mask(U, X, C)                              \
 638   ((__mmask32) __builtin_ia32_fpclassbf16512_mask (                        \
 639       (__v32bf) (__m512bh) (X), (int) (C), (__mmask32) (U)))
 640
 641 #define _mm512_fpclass_pbh_mask(X, C)                                      \
 642   ((__mmask32) __builtin_ia32_fpclassbf16512_mask (                        \
 643       (__v32bf) (__m512bh) (X), (int) (C), (__mmask32) (-1)))
 644 #endif /* __OPIMTIZE__ */
 645
 646
 647 /* Intrinsics vcmpbf16.  */
 648 #ifdef __OPTIMIZE__
 649 extern __inline __mmask32
 650 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 651 _mm512_mask_cmp_pbh_mask (__mmask32 __U, __m512bh __A, __m512bh __B,
 652                           const int __imm)
 653 {
 654   return (__mmask32)
 655     __builtin_ia32_cmpbf16512_mask (__A, __B, __imm, __U);
 656 }
 657
 658 extern __inline __mmask32
 659 __attribute__ ((__gnu_inline__, __always_inline__, __artificial__))
 660 _mm512_cmp_pbh_mask (__m512bh __A, __m512bh __B, const int __imm)
 661 {
 662   return (__mmask32)
 663     __builtin_ia32_cmpbf16512_mask (__A, __B, __imm,
 664                                     (__mmask32) -1);
 665 }
 666
 667 #else
 668 #define _mm512_mask_cmp_pbh_mask(A, B, C, D)                            \
 669   ((__mmask32) __builtin_ia32_cmpbf16512_mask ((B), (C), (D), (A)))
 670
 671 #define _mm512_cmp_pbh_mask(A, B, C)                                    \
 672   ((__mmask32) __builtin_ia32_cmpbf16512_mask ((A), (B), (C), (-1)))
 673
 674 #endif /* __OPIMTIZE__ */
 675
 676 #ifdef __DISABLE_AVX10_2_512__
 677 #undef __DISABLE_AVX10_2_512__
 678 #pragma GCC pop_options
 679 #endif /* __DISABLE_AVX10_2_512__ */
 680
 681 #endif /* _AVX10_2_512BF16INTRIN_H_INCLUDED */