| 1 |
/* Copyright (C) 2013-2021 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 <avx512vbmi2intrin.h> directly; include <immintrin.h> instead." |
| 26 |
#endif |
| 27 |
|
| 28 |
#ifndef __AVX512VBMI2INTRIN_H_INCLUDED |
| 29 |
#define __AVX512VBMI2INTRIN_H_INCLUDED |
| 30 |
|
| 31 |
#if !defined(__AVX512VBMI2__) |
| 32 |
#pragma GCC push_options |
| 33 |
#pragma GCC target("avx512vbmi2") |
| 34 |
#define __DISABLE_AVX512VBMI2__ |
| 35 |
#endif /* __AVX512VBMI2__ */ |
| 36 |
|
| 37 |
#ifdef __OPTIMIZE__ |
| 38 |
extern __inline __m512i |
| 39 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 40 |
_mm512_shrdi_epi16 (__m512i __A, __m512i __B, int __C) |
| 41 |
{ |
| 42 |
return (__m512i) __builtin_ia32_vpshrd_v32hi ((__v32hi)__A, (__v32hi) __B, |
| 43 |
__C); |
| 44 |
} |
| 45 |
|
| 46 |
extern __inline __m512i |
| 47 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 48 |
_mm512_shrdi_epi32 (__m512i __A, __m512i __B, int __C) |
| 49 |
{ |
| 50 |
return (__m512i) __builtin_ia32_vpshrd_v16si ((__v16si)__A, (__v16si) __B, |
| 51 |
__C); |
| 52 |
} |
| 53 |
|
| 54 |
extern __inline __m512i |
| 55 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 56 |
_mm512_mask_shrdi_epi32 (__m512i __A, __mmask16 __B, __m512i __C, __m512i __D, |
| 57 |
int __E) |
| 58 |
{ |
| 59 |
return (__m512i)__builtin_ia32_vpshrd_v16si_mask ((__v16si)__C, |
| 60 |
(__v16si) __D, __E, (__v16si) __A, (__mmask16)__B); |
| 61 |
} |
| 62 |
|
| 63 |
extern __inline __m512i |
| 64 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 65 |
_mm512_maskz_shrdi_epi32 (__mmask16 __A, __m512i __B, __m512i __C, int __D) |
| 66 |
{ |
| 67 |
return (__m512i)__builtin_ia32_vpshrd_v16si_mask ((__v16si)__B, |
| 68 |
(__v16si) __C, __D, (__v16si) _mm512_setzero_si512 (), (__mmask16)__A); |
| 69 |
} |
| 70 |
|
| 71 |
extern __inline __m512i |
| 72 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 73 |
_mm512_shrdi_epi64 (__m512i __A, __m512i __B, int __C) |
| 74 |
{ |
| 75 |
return (__m512i) __builtin_ia32_vpshrd_v8di ((__v8di)__A, (__v8di) __B, __C); |
| 76 |
} |
| 77 |
|
| 78 |
extern __inline __m512i |
| 79 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 80 |
_mm512_mask_shrdi_epi64 (__m512i __A, __mmask8 __B, __m512i __C, __m512i __D, |
| 81 |
int __E) |
| 82 |
{ |
| 83 |
return (__m512i)__builtin_ia32_vpshrd_v8di_mask ((__v8di)__C, (__v8di) __D, |
| 84 |
__E, (__v8di) __A, (__mmask8)__B); |
| 85 |
} |
| 86 |
|
| 87 |
extern __inline __m512i |
| 88 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 89 |
_mm512_maskz_shrdi_epi64 (__mmask8 __A, __m512i __B, __m512i __C, int __D) |
| 90 |
{ |
| 91 |
return (__m512i)__builtin_ia32_vpshrd_v8di_mask ((__v8di)__B, (__v8di) __C, |
| 92 |
__D, (__v8di) _mm512_setzero_si512 (), (__mmask8)__A); |
| 93 |
} |
| 94 |
|
| 95 |
extern __inline __m512i |
| 96 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 97 |
_mm512_shldi_epi16 (__m512i __A, __m512i __B, int __C) |
| 98 |
{ |
| 99 |
return (__m512i) __builtin_ia32_vpshld_v32hi ((__v32hi)__A, (__v32hi) __B, |
| 100 |
__C); |
| 101 |
} |
| 102 |
|
| 103 |
extern __inline __m512i |
| 104 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 105 |
_mm512_shldi_epi32 (__m512i __A, __m512i __B, int __C) |
| 106 |
{ |
| 107 |
return (__m512i) __builtin_ia32_vpshld_v16si ((__v16si)__A, (__v16si) __B, |
| 108 |
__C); |
| 109 |
} |
| 110 |
|
| 111 |
extern __inline __m512i |
| 112 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 113 |
_mm512_mask_shldi_epi32 (__m512i __A, __mmask16 __B, __m512i __C, __m512i __D, |
| 114 |
int __E) |
| 115 |
{ |
| 116 |
return (__m512i)__builtin_ia32_vpshld_v16si_mask ((__v16si)__C, |
| 117 |
(__v16si) __D, __E, (__v16si) __A, (__mmask16)__B); |
| 118 |
} |
| 119 |
|
| 120 |
extern __inline __m512i |
| 121 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 122 |
_mm512_maskz_shldi_epi32 (__mmask16 __A, __m512i __B, __m512i __C, int __D) |
| 123 |
{ |
| 124 |
return (__m512i)__builtin_ia32_vpshld_v16si_mask ((__v16si)__B, |
| 125 |
(__v16si) __C, __D, (__v16si) _mm512_setzero_si512 (), (__mmask16)__A); |
| 126 |
} |
| 127 |
|
| 128 |
extern __inline __m512i |
| 129 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 130 |
_mm512_shldi_epi64 (__m512i __A, __m512i __B, int __C) |
| 131 |
{ |
| 132 |
return (__m512i) __builtin_ia32_vpshld_v8di ((__v8di)__A, (__v8di) __B, __C); |
| 133 |
} |
| 134 |
|
| 135 |
extern __inline __m512i |
| 136 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 137 |
_mm512_mask_shldi_epi64 (__m512i __A, __mmask8 __B, __m512i __C, __m512i __D, |
| 138 |
int __E) |
| 139 |
{ |
| 140 |
return (__m512i)__builtin_ia32_vpshld_v8di_mask ((__v8di)__C, (__v8di) __D, |
| 141 |
__E, (__v8di) __A, (__mmask8)__B); |
| 142 |
} |
| 143 |
|
| 144 |
extern __inline __m512i |
| 145 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 146 |
_mm512_maskz_shldi_epi64 (__mmask8 __A, __m512i __B, __m512i __C, int __D) |
| 147 |
{ |
| 148 |
return (__m512i)__builtin_ia32_vpshld_v8di_mask ((__v8di)__B, (__v8di) __C, |
| 149 |
__D, (__v8di) _mm512_setzero_si512 (), (__mmask8)__A); |
| 150 |
} |
| 151 |
#else |
| 152 |
#define _mm512_shrdi_epi16(A, B, C) \ |
| 153 |
((__m512i) __builtin_ia32_vpshrd_v32hi ((__v32hi)(__m512i)(A), \ |
| 154 |
(__v32hi)(__m512i)(B),(int)(C))) |
| 155 |
#define _mm512_shrdi_epi32(A, B, C) \ |
| 156 |
((__m512i) __builtin_ia32_vpshrd_v16si ((__v16si)(__m512i)(A), \ |
| 157 |
(__v16si)(__m512i)(B),(int)(C))) |
| 158 |
#define _mm512_mask_shrdi_epi32(A, B, C, D, E) \ |
| 159 |
((__m512i) __builtin_ia32_vpshrd_v16si_mask ((__v16si)(__m512i)(C), \ |
| 160 |
(__v16si)(__m512i)(D), \ |
| 161 |
(int)(E), \ |
| 162 |
(__v16si)(__m512i)(A), \ |
| 163 |
(__mmask16)(B))) |
| 164 |
#define _mm512_maskz_shrdi_epi32(A, B, C, D) \ |
| 165 |
((__m512i) \ |
| 166 |
__builtin_ia32_vpshrd_v16si_mask ((__v16si)(__m512i)(B), \ |
| 167 |
(__v16si)(__m512i)(C),(int)(D), \ |
| 168 |
(__v16si)(__m512i)_mm512_setzero_si512 (), \ |
| 169 |
(__mmask16)(A))) |
| 170 |
#define _mm512_shrdi_epi64(A, B, C) \ |
| 171 |
((__m512i) __builtin_ia32_vpshrd_v8di ((__v8di)(__m512i)(A), \ |
| 172 |
(__v8di)(__m512i)(B),(int)(C))) |
| 173 |
#define _mm512_mask_shrdi_epi64(A, B, C, D, E) \ |
| 174 |
((__m512i) __builtin_ia32_vpshrd_v8di_mask ((__v8di)(__m512i)(C), \ |
| 175 |
(__v8di)(__m512i)(D), (int)(E), \ |
| 176 |
(__v8di)(__m512i)(A), \ |
| 177 |
(__mmask8)(B))) |
| 178 |
#define _mm512_maskz_shrdi_epi64(A, B, C, D) \ |
| 179 |
((__m512i) \ |
| 180 |
__builtin_ia32_vpshrd_v8di_mask ((__v8di)(__m512i)(B), \ |
| 181 |
(__v8di)(__m512i)(C),(int)(D), \ |
| 182 |
(__v8di)(__m512i)_mm512_setzero_si512 (), \ |
| 183 |
(__mmask8)(A))) |
| 184 |
#define _mm512_shldi_epi16(A, B, C) \ |
| 185 |
((__m512i) __builtin_ia32_vpshld_v32hi ((__v32hi)(__m512i)(A), \ |
| 186 |
(__v32hi)(__m512i)(B),(int)(C))) |
| 187 |
#define _mm512_shldi_epi32(A, B, C) \ |
| 188 |
((__m512i) __builtin_ia32_vpshld_v16si ((__v16si)(__m512i)(A), \ |
| 189 |
(__v16si)(__m512i)(B),(int)(C))) |
| 190 |
#define _mm512_mask_shldi_epi32(A, B, C, D, E) \ |
| 191 |
((__m512i) __builtin_ia32_vpshld_v16si_mask ((__v16si)(__m512i)(C), \ |
| 192 |
(__v16si)(__m512i)(D), \ |
| 193 |
(int)(E), \ |
| 194 |
(__v16si)(__m512i)(A), \ |
| 195 |
(__mmask16)(B))) |
| 196 |
#define _mm512_maskz_shldi_epi32(A, B, C, D) \ |
| 197 |
((__m512i) \ |
| 198 |
__builtin_ia32_vpshld_v16si_mask ((__v16si)(__m512i)(B), \ |
| 199 |
(__v16si)(__m512i)(C),(int)(D), \ |
| 200 |
(__v16si)(__m512i)_mm512_setzero_si512 (), \ |
| 201 |
(__mmask16)(A))) |
| 202 |
#define _mm512_shldi_epi64(A, B, C) \ |
| 203 |
((__m512i) __builtin_ia32_vpshld_v8di ((__v8di)(__m512i)(A), \ |
| 204 |
(__v8di)(__m512i)(B), (int)(C))) |
| 205 |
#define _mm512_mask_shldi_epi64(A, B, C, D, E) \ |
| 206 |
((__m512i) __builtin_ia32_vpshld_v8di_mask ((__v8di)(__m512i)(C), \ |
| 207 |
(__v8di)(__m512i)(D), (int)(E), \ |
| 208 |
(__v8di)(__m512i)(A), \ |
| 209 |
(__mmask8)(B))) |
| 210 |
#define _mm512_maskz_shldi_epi64(A, B, C, D) \ |
| 211 |
((__m512i) \ |
| 212 |
__builtin_ia32_vpshld_v8di_mask ((__v8di)(__m512i)(B), \ |
| 213 |
(__v8di)(__m512i)(C),(int)(D), \ |
| 214 |
(__v8di)(__m512i)_mm512_setzero_si512 (), \ |
| 215 |
(__mmask8)(A))) |
| 216 |
#endif |
| 217 |
|
| 218 |
extern __inline __m512i |
| 219 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 220 |
_mm512_shrdv_epi16 (__m512i __A, __m512i __B, __m512i __C) |
| 221 |
{ |
| 222 |
return (__m512i) __builtin_ia32_vpshrdv_v32hi ((__v32hi)__A, (__v32hi) __B, |
| 223 |
(__v32hi) __C); |
| 224 |
} |
| 225 |
|
| 226 |
extern __inline __m512i |
| 227 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 228 |
_mm512_shrdv_epi32 (__m512i __A, __m512i __B, __m512i __C) |
| 229 |
{ |
| 230 |
return (__m512i) __builtin_ia32_vpshrdv_v16si ((__v16si)__A, (__v16si) __B, |
| 231 |
(__v16si) __C); |
| 232 |
} |
| 233 |
|
| 234 |
extern __inline __m512i |
| 235 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 236 |
_mm512_mask_shrdv_epi32 (__m512i __A, __mmask16 __B, __m512i __C, __m512i __D) |
| 237 |
{ |
| 238 |
return (__m512i)__builtin_ia32_vpshrdv_v16si_mask ((__v16si)__A, |
| 239 |
(__v16si) __C, (__v16si) __D, (__mmask16)__B); |
| 240 |
} |
| 241 |
|
| 242 |
extern __inline __m512i |
| 243 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 244 |
_mm512_maskz_shrdv_epi32 (__mmask16 __A, __m512i __B, __m512i __C, __m512i __D) |
| 245 |
{ |
| 246 |
return (__m512i)__builtin_ia32_vpshrdv_v16si_maskz ((__v16si)__B, |
| 247 |
(__v16si) __C, (__v16si) __D, (__mmask16)__A); |
| 248 |
} |
| 249 |
|
| 250 |
extern __inline __m512i |
| 251 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 252 |
_mm512_shrdv_epi64 (__m512i __A, __m512i __B, __m512i __C) |
| 253 |
{ |
| 254 |
return (__m512i) __builtin_ia32_vpshrdv_v8di ((__v8di)__A, (__v8di) __B, |
| 255 |
(__v8di) __C); |
| 256 |
} |
| 257 |
|
| 258 |
extern __inline __m512i |
| 259 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 260 |
_mm512_mask_shrdv_epi64 (__m512i __A, __mmask8 __B, __m512i __C, __m512i __D) |
| 261 |
{ |
| 262 |
return (__m512i)__builtin_ia32_vpshrdv_v8di_mask ((__v8di)__A, (__v8di) __C, |
| 263 |
(__v8di) __D, (__mmask8)__B); |
| 264 |
} |
| 265 |
|
| 266 |
extern __inline __m512i |
| 267 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 268 |
_mm512_maskz_shrdv_epi64 (__mmask8 __A, __m512i __B, __m512i __C, __m512i __D) |
| 269 |
{ |
| 270 |
return (__m512i)__builtin_ia32_vpshrdv_v8di_maskz ((__v8di)__B, (__v8di) __C, |
| 271 |
(__v8di) __D, (__mmask8)__A); |
| 272 |
} |
| 273 |
extern __inline __m512i |
| 274 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 275 |
_mm512_shldv_epi16 (__m512i __A, __m512i __B, __m512i __C) |
| 276 |
{ |
| 277 |
return (__m512i) __builtin_ia32_vpshldv_v32hi ((__v32hi)__A, (__v32hi) __B, |
| 278 |
(__v32hi) __C); |
| 279 |
} |
| 280 |
|
| 281 |
extern __inline __m512i |
| 282 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 283 |
_mm512_shldv_epi32 (__m512i __A, __m512i __B, __m512i __C) |
| 284 |
{ |
| 285 |
return (__m512i) __builtin_ia32_vpshldv_v16si ((__v16si)__A, (__v16si) __B, |
| 286 |
(__v16si) __C); |
| 287 |
} |
| 288 |
|
| 289 |
extern __inline __m512i |
| 290 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 291 |
_mm512_mask_shldv_epi32 (__m512i __A, __mmask16 __B, __m512i __C, __m512i __D) |
| 292 |
{ |
| 293 |
return (__m512i)__builtin_ia32_vpshldv_v16si_mask ((__v16si)__A, |
| 294 |
(__v16si) __C, (__v16si) __D, (__mmask16)__B); |
| 295 |
} |
| 296 |
|
| 297 |
extern __inline __m512i |
| 298 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 299 |
_mm512_maskz_shldv_epi32 (__mmask16 __A, __m512i __B, __m512i __C, __m512i __D) |
| 300 |
{ |
| 301 |
return (__m512i)__builtin_ia32_vpshldv_v16si_maskz ((__v16si)__B, |
| 302 |
(__v16si) __C, (__v16si) __D, (__mmask16)__A); |
| 303 |
} |
| 304 |
|
| 305 |
extern __inline __m512i |
| 306 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 307 |
_mm512_shldv_epi64 (__m512i __A, __m512i __B, __m512i __C) |
| 308 |
{ |
| 309 |
return (__m512i) __builtin_ia32_vpshldv_v8di ((__v8di)__A, (__v8di) __B, |
| 310 |
(__v8di) __C); |
| 311 |
} |
| 312 |
|
| 313 |
extern __inline __m512i |
| 314 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 315 |
_mm512_mask_shldv_epi64 (__m512i __A, __mmask8 __B, __m512i __C, __m512i __D) |
| 316 |
{ |
| 317 |
return (__m512i)__builtin_ia32_vpshldv_v8di_mask ((__v8di)__A, (__v8di) __C, |
| 318 |
(__v8di) __D, (__mmask8)__B); |
| 319 |
} |
| 320 |
|
| 321 |
extern __inline __m512i |
| 322 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 323 |
_mm512_maskz_shldv_epi64 (__mmask8 __A, __m512i __B, __m512i __C, __m512i __D) |
| 324 |
{ |
| 325 |
return (__m512i)__builtin_ia32_vpshldv_v8di_maskz ((__v8di)__B, (__v8di) __C, |
| 326 |
(__v8di) __D, (__mmask8)__A); |
| 327 |
} |
| 328 |
|
| 329 |
#ifdef __DISABLE_AVX512VBMI2__ |
| 330 |
#undef __DISABLE_AVX512VBMI2__ |
| 331 |
|
| 332 |
#pragma GCC pop_options |
| 333 |
#endif /* __DISABLE_AVX512VBMI2__ */ |
| 334 |
|
| 335 |
#if !defined(__AVX512VBMI2__) || !defined(__AVX512BW__) |
| 336 |
#pragma GCC push_options |
| 337 |
#pragma GCC target("avx512vbmi2,avx512bw") |
| 338 |
#define __DISABLE_AVX512VBMI2BW__ |
| 339 |
#endif /* __AVX512VBMI2BW__ */ |
| 340 |
|
| 341 |
extern __inline __m512i |
| 342 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 343 |
_mm512_mask_compress_epi8 (__m512i __A, __mmask64 __B, __m512i __C) |
| 344 |
{ |
| 345 |
return (__m512i) __builtin_ia32_compressqi512_mask ((__v64qi)__C, |
| 346 |
(__v64qi)__A, (__mmask64)__B); |
| 347 |
} |
| 348 |
|
| 349 |
|
| 350 |
extern __inline __m512i |
| 351 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 352 |
_mm512_maskz_compress_epi8 (__mmask64 __A, __m512i __B) |
| 353 |
{ |
| 354 |
return (__m512i) __builtin_ia32_compressqi512_mask ((__v64qi)__B, |
| 355 |
(__v64qi)_mm512_setzero_si512 (), (__mmask64)__A); |
| 356 |
} |
| 357 |
|
| 358 |
|
| 359 |
extern __inline void |
| 360 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 361 |
_mm512_mask_compressstoreu_epi8 (void * __A, __mmask64 __B, __m512i __C) |
| 362 |
{ |
| 363 |
__builtin_ia32_compressstoreuqi512_mask ((__v64qi *) __A, (__v64qi) __C, |
| 364 |
(__mmask64) __B); |
| 365 |
} |
| 366 |
|
| 367 |
extern __inline __m512i |
| 368 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 369 |
_mm512_mask_compress_epi16 (__m512i __A, __mmask32 __B, __m512i __C) |
| 370 |
{ |
| 371 |
return (__m512i) __builtin_ia32_compresshi512_mask ((__v32hi)__C, |
| 372 |
(__v32hi)__A, (__mmask32)__B); |
| 373 |
} |
| 374 |
|
| 375 |
extern __inline __m512i |
| 376 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 377 |
_mm512_maskz_compress_epi16 (__mmask32 __A, __m512i __B) |
| 378 |
{ |
| 379 |
return (__m512i) __builtin_ia32_compresshi512_mask ((__v32hi)__B, |
| 380 |
(__v32hi)_mm512_setzero_si512 (), (__mmask32)__A); |
| 381 |
} |
| 382 |
|
| 383 |
extern __inline void |
| 384 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 385 |
_mm512_mask_compressstoreu_epi16 (void * __A, __mmask32 __B, __m512i __C) |
| 386 |
{ |
| 387 |
__builtin_ia32_compressstoreuhi512_mask ((__v32hi *) __A, (__v32hi) __C, |
| 388 |
(__mmask32) __B); |
| 389 |
} |
| 390 |
|
| 391 |
extern __inline __m512i |
| 392 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 393 |
_mm512_mask_expand_epi8 (__m512i __A, __mmask64 __B, __m512i __C) |
| 394 |
{ |
| 395 |
return (__m512i) __builtin_ia32_expandqi512_mask ((__v64qi) __C, |
| 396 |
(__v64qi) __A, |
| 397 |
(__mmask64) __B); |
| 398 |
} |
| 399 |
|
| 400 |
extern __inline __m512i |
| 401 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 402 |
_mm512_maskz_expand_epi8 (__mmask64 __A, __m512i __B) |
| 403 |
{ |
| 404 |
return (__m512i) __builtin_ia32_expandqi512_maskz ((__v64qi) __B, |
| 405 |
(__v64qi) _mm512_setzero_si512 (), (__mmask64) __A); |
| 406 |
} |
| 407 |
|
| 408 |
extern __inline __m512i |
| 409 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 410 |
_mm512_mask_expandloadu_epi8 (__m512i __A, __mmask64 __B, const void * __C) |
| 411 |
{ |
| 412 |
return (__m512i) __builtin_ia32_expandloadqi512_mask ((const __v64qi *) __C, |
| 413 |
(__v64qi) __A, (__mmask64) __B); |
| 414 |
} |
| 415 |
|
| 416 |
extern __inline __m512i |
| 417 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 418 |
_mm512_maskz_expandloadu_epi8 (__mmask64 __A, const void * __B) |
| 419 |
{ |
| 420 |
return (__m512i) __builtin_ia32_expandloadqi512_maskz ((const __v64qi *) __B, |
| 421 |
(__v64qi) _mm512_setzero_si512 (), (__mmask64) __A); |
| 422 |
} |
| 423 |
|
| 424 |
extern __inline __m512i |
| 425 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 426 |
_mm512_mask_expand_epi16 (__m512i __A, __mmask32 __B, __m512i __C) |
| 427 |
{ |
| 428 |
return (__m512i) __builtin_ia32_expandhi512_mask ((__v32hi) __C, |
| 429 |
(__v32hi) __A, |
| 430 |
(__mmask32) __B); |
| 431 |
} |
| 432 |
|
| 433 |
extern __inline __m512i |
| 434 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 435 |
_mm512_maskz_expand_epi16 (__mmask32 __A, __m512i __B) |
| 436 |
{ |
| 437 |
return (__m512i) __builtin_ia32_expandhi512_maskz ((__v32hi) __B, |
| 438 |
(__v32hi) _mm512_setzero_si512 (), (__mmask32) __A); |
| 439 |
} |
| 440 |
|
| 441 |
extern __inline __m512i |
| 442 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 443 |
_mm512_mask_expandloadu_epi16 (__m512i __A, __mmask32 __B, const void * __C) |
| 444 |
{ |
| 445 |
return (__m512i) __builtin_ia32_expandloadhi512_mask ((const __v32hi *) __C, |
| 446 |
(__v32hi) __A, (__mmask32) __B); |
| 447 |
} |
| 448 |
|
| 449 |
extern __inline __m512i |
| 450 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 451 |
_mm512_maskz_expandloadu_epi16 (__mmask32 __A, const void * __B) |
| 452 |
{ |
| 453 |
return (__m512i) __builtin_ia32_expandloadhi512_maskz ((const __v32hi *) __B, |
| 454 |
(__v32hi) _mm512_setzero_si512 (), (__mmask32) __A); |
| 455 |
} |
| 456 |
|
| 457 |
#ifdef __OPTIMIZE__ |
| 458 |
extern __inline __m512i |
| 459 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 460 |
_mm512_mask_shrdi_epi16 (__m512i __A, __mmask32 __B, __m512i __C, __m512i __D, |
| 461 |
int __E) |
| 462 |
{ |
| 463 |
return (__m512i)__builtin_ia32_vpshrd_v32hi_mask ((__v32hi)__C, |
| 464 |
(__v32hi) __D, __E, (__v32hi) __A, (__mmask32)__B); |
| 465 |
} |
| 466 |
|
| 467 |
extern __inline __m512i |
| 468 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 469 |
_mm512_maskz_shrdi_epi16 (__mmask32 __A, __m512i __B, __m512i __C, int __D) |
| 470 |
{ |
| 471 |
return (__m512i)__builtin_ia32_vpshrd_v32hi_mask ((__v32hi)__B, |
| 472 |
(__v32hi) __C, __D, (__v32hi) _mm512_setzero_si512 (), (__mmask32)__A); |
| 473 |
} |
| 474 |
|
| 475 |
extern __inline __m512i |
| 476 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 477 |
_mm512_mask_shldi_epi16 (__m512i __A, __mmask32 __B, __m512i __C, __m512i __D, |
| 478 |
int __E) |
| 479 |
{ |
| 480 |
return (__m512i)__builtin_ia32_vpshld_v32hi_mask ((__v32hi)__C, |
| 481 |
(__v32hi) __D, __E, (__v32hi) __A, (__mmask32)__B); |
| 482 |
} |
| 483 |
|
| 484 |
extern __inline __m512i |
| 485 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 486 |
_mm512_maskz_shldi_epi16 (__mmask32 __A, __m512i __B, __m512i __C, int __D) |
| 487 |
{ |
| 488 |
return (__m512i)__builtin_ia32_vpshld_v32hi_mask ((__v32hi)__B, |
| 489 |
(__v32hi) __C, __D, (__v32hi) _mm512_setzero_si512 (), (__mmask32)__A); |
| 490 |
} |
| 491 |
|
| 492 |
#else |
| 493 |
#define _mm512_mask_shrdi_epi16(A, B, C, D, E) \ |
| 494 |
((__m512i) __builtin_ia32_vpshrd_v32hi_mask ((__v32hi)(__m512i)(C), \ |
| 495 |
(__v32hi)(__m512i)(D), \ |
| 496 |
(int)(E), \ |
| 497 |
(__v32hi)(__m512i)(A), \ |
| 498 |
(__mmask32)(B))) |
| 499 |
#define _mm512_maskz_shrdi_epi16(A, B, C, D) \ |
| 500 |
((__m512i) \ |
| 501 |
__builtin_ia32_vpshrd_v32hi_mask ((__v32hi)(__m512i)(B), \ |
| 502 |
(__v32hi)(__m512i)(C),(int)(D), \ |
| 503 |
(__v32hi)(__m512i)_mm512_setzero_si512 (), \ |
| 504 |
(__mmask32)(A))) |
| 505 |
#define _mm512_mask_shldi_epi16(A, B, C, D, E) \ |
| 506 |
((__m512i) __builtin_ia32_vpshld_v32hi_mask ((__v32hi)(__m512i)(C), \ |
| 507 |
(__v32hi)(__m512i)(D), \ |
| 508 |
(int)(E), \ |
| 509 |
(__v32hi)(__m512i)(A), \ |
| 510 |
(__mmask32)(B))) |
| 511 |
#define _mm512_maskz_shldi_epi16(A, B, C, D) \ |
| 512 |
((__m512i) \ |
| 513 |
__builtin_ia32_vpshld_v32hi_mask ((__v32hi)(__m512i)(B), \ |
| 514 |
(__v32hi)(__m512i)(C),(int)(D), \ |
| 515 |
(__v32hi)(__m512i)_mm512_setzero_si512 (), \ |
| 516 |
(__mmask32)(A))) |
| 517 |
#endif |
| 518 |
|
| 519 |
extern __inline __m512i |
| 520 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 521 |
_mm512_mask_shrdv_epi16 (__m512i __A, __mmask32 __B, __m512i __C, __m512i __D) |
| 522 |
{ |
| 523 |
return (__m512i)__builtin_ia32_vpshrdv_v32hi_mask ((__v32hi)__A, |
| 524 |
(__v32hi) __C, (__v32hi) __D, (__mmask32)__B); |
| 525 |
} |
| 526 |
|
| 527 |
extern __inline __m512i |
| 528 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 529 |
_mm512_maskz_shrdv_epi16 (__mmask32 __A, __m512i __B, __m512i __C, __m512i __D) |
| 530 |
{ |
| 531 |
return (__m512i)__builtin_ia32_vpshrdv_v32hi_maskz ((__v32hi)__B, |
| 532 |
(__v32hi) __C, (__v32hi) __D, (__mmask32)__A); |
| 533 |
} |
| 534 |
|
| 535 |
extern __inline __m512i |
| 536 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 537 |
_mm512_mask_shldv_epi16 (__m512i __A, __mmask32 __B, __m512i __C, __m512i __D) |
| 538 |
{ |
| 539 |
return (__m512i)__builtin_ia32_vpshldv_v32hi_mask ((__v32hi)__A, |
| 540 |
(__v32hi) __C, (__v32hi) __D, (__mmask32)__B); |
| 541 |
} |
| 542 |
|
| 543 |
extern __inline __m512i |
| 544 |
__attribute__((__gnu_inline__, __always_inline__, __artificial__)) |
| 545 |
_mm512_maskz_shldv_epi16 (__mmask32 __A, __m512i __B, __m512i __C, __m512i __D) |
| 546 |
{ |
| 547 |
return (__m512i)__builtin_ia32_vpshldv_v32hi_maskz ((__v32hi)__B, |
| 548 |
(__v32hi) __C, (__v32hi) __D, (__mmask32)__A); |
| 549 |
} |
| 550 |
|
| 551 |
#ifdef __DISABLE_AVX512VBMI2BW__ |
| 552 |
#undef __DISABLE_AVX512VBMI2BW__ |
| 553 |
|
| 554 |
#pragma GCC pop_options |
| 555 |
#endif /* __DISABLE_AVX512VBMI2BW__ */ |
| 556 |
|
| 557 |
#endif /* __AVX512VBMI2INTRIN_H_INCLUDED */ |
