| 1 | /*===----------- avx10_2_512bf16intrin.h - AVX10-BF16 intrinsics ---------=== |
| 2 | * |
| 3 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | * See https://llvm.org/LICENSE.txt for license information. |
| 5 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | * |
| 7 | *===-----------------------------------------------------------------------=== |
| 8 | */ |
| 9 | #ifndef __IMMINTRIN_H |
| 10 | #error \ |
| 11 | "Never use <avx10_2_512bf16intrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif |
| 13 | |
| 14 | #ifdef __SSE2__ |
| 15 | |
| 16 | #ifndef __AVX10_2_512BF16INTRIN_H |
| 17 | #define __AVX10_2_512BF16INTRIN_H |
| 18 | |
| 19 | /* Define the default attributes for the functions in this file. */ |
| 20 | typedef __bf16 __m512bh_u __attribute__((__vector_size__(64), __aligned__(1))); |
| 21 | |
| 22 | /* Define the default attributes for the functions in this file. */ |
| 23 | #define __DEFAULT_FN_ATTRS512 \ |
| 24 | __attribute__((__always_inline__, __nodebug__, __target__("avx10.2"), \ |
| 25 | __min_vector_width__(512))) |
| 26 | |
| 27 | #if defined(__cplusplus) && (__cplusplus >= 201103L) |
| 28 | #define __DEFAULT_FN_ATTRS512_CONSTEXPR __DEFAULT_FN_ATTRS512 constexpr |
| 29 | #else |
| 30 | #define __DEFAULT_FN_ATTRS512_CONSTEXPR __DEFAULT_FN_ATTRS512 |
| 31 | #endif |
| 32 | |
| 33 | static __inline __m512bh __DEFAULT_FN_ATTRS512 _mm512_setzero_pbh(void) { |
| 34 | return __builtin_bit_cast(__m512bh, _mm512_setzero_ps()); |
| 35 | } |
| 36 | |
| 37 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_undefined_pbh(void) { |
| 38 | return (__m512bh)__builtin_ia32_undef512(); |
| 39 | } |
| 40 | |
| 41 | static __inline __m512bh __DEFAULT_FN_ATTRS512 _mm512_set1_pbh(__bf16 bf) { |
| 42 | return (__m512bh)(__v32bf){bf, bf, bf, bf, bf, bf, bf, bf, bf, bf, bf, |
| 43 | bf, bf, bf, bf, bf, bf, bf, bf, bf, bf, bf, |
| 44 | bf, bf, bf, bf, bf, bf, bf, bf, bf, bf}; |
| 45 | } |
| 46 | |
| 47 | static __inline __m512bh __DEFAULT_FN_ATTRS512 _mm512_set_pbh( |
| 48 | __bf16 bf1, __bf16 bf2, __bf16 bf3, __bf16 bf4, __bf16 bf5, __bf16 bf6, |
| 49 | __bf16 bf7, __bf16 bf8, __bf16 bf9, __bf16 bf10, __bf16 bf11, __bf16 bf12, |
| 50 | __bf16 bf13, __bf16 bf14, __bf16 bf15, __bf16 bf16, __bf16 bf17, |
| 51 | __bf16 bf18, __bf16 bf19, __bf16 bf20, __bf16 bf21, __bf16 bf22, |
| 52 | __bf16 bf23, __bf16 bf24, __bf16 bf25, __bf16 bf26, __bf16 bf27, |
| 53 | __bf16 bf28, __bf16 bf29, __bf16 bf30, __bf16 bf31, __bf16 bf32) { |
| 54 | return (__m512bh)(__v32bf){bf32, bf31, bf30, bf29, bf28, bf27, bf26, bf25, |
| 55 | bf24, bf23, bf22, bf21, bf20, bf19, bf18, bf17, |
| 56 | bf16, bf15, bf14, bf13, bf12, bf11, bf10, bf9, |
| 57 | bf8, bf7, bf6, bf5, bf4, bf3, bf2, bf1}; |
| 58 | } |
| 59 | |
| 60 | #define _mm512_setr_pbh(bf1, bf2, bf3, bf4, bf5, bf6, bf7, bf8, bf9, bf10, \ |
| 61 | bf11, bf12, bf13, bf14, bf15, bf16, bf17, bf18, bf19, \ |
| 62 | bf20, bf21, bf22, bf23, bf24, bf25, bf26, bf27, bf28, \ |
| 63 | bf29, bf30, bf31, bf32) \ |
| 64 | _mm512_set_pbh((bf32), (bf31), (bf30), (bf29), (bf28), (bf27), (bf26), \ |
| 65 | (bf25), (bf24), (bf23), (bf22), (bf21), (bf20), (bf19), \ |
| 66 | (bf18), (bf17), (bf16), (bf15), (bf14), (bf13), (bf12), \ |
| 67 | (bf11), (bf10), (bf9), (bf8), (bf7), (bf6), (bf5), (bf4), \ |
| 68 | (bf3), (bf2), (bf1)) |
| 69 | |
| 70 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 71 | _mm512_castbf16_ps(__m512bh __a) { |
| 72 | return (__m512)__a; |
| 73 | } |
| 74 | |
| 75 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 76 | _mm512_castbf16_pd(__m512bh __a) { |
| 77 | return (__m512d)__a; |
| 78 | } |
| 79 | |
| 80 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 81 | _mm512_castbf16_si512(__m512bh __a) { |
| 82 | return (__m512i)__a; |
| 83 | } |
| 84 | |
| 85 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_castps_pbh(__m512 __a) { |
| 86 | return (__m512bh)__a; |
| 87 | } |
| 88 | |
| 89 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 90 | _mm512_castpd_pbh(__m512d __a) { |
| 91 | return (__m512bh)__a; |
| 92 | } |
| 93 | |
| 94 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 95 | _mm512_castsi512_pbh(__m512i __a) { |
| 96 | return (__m512bh)__a; |
| 97 | } |
| 98 | |
| 99 | static __inline__ __m128bh __DEFAULT_FN_ATTRS512 |
| 100 | _mm512_castbf16512_pbh128(__m512bh __a) { |
| 101 | return __builtin_shufflevector(__a, __a, 0, 1, 2, 3, 4, 5, 6, 7); |
| 102 | } |
| 103 | |
| 104 | static __inline__ __m256bh __DEFAULT_FN_ATTRS512 |
| 105 | _mm512_castbf16512_pbh256(__m512bh __a) { |
| 106 | return __builtin_shufflevector(__a, __a, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, |
| 107 | 12, 13, 14, 15); |
| 108 | } |
| 109 | |
| 110 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 111 | _mm512_castbf16128_pbh512(__m128bh __a) { |
| 112 | return __builtin_shufflevector(__a, __a, 0, 1, 2, 3, 4, 5, 6, 7, -1, -1, -1, |
| 113 | -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, |
| 114 | -1, -1, -1, -1, -1, -1, -1, -1, -1); |
| 115 | } |
| 116 | |
| 117 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 118 | _mm512_castbf16256_pbh512(__m256bh __a) { |
| 119 | return __builtin_shufflevector(__a, __a, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, |
| 120 | 12, 13, 14, 15, -1, -1, -1, -1, -1, -1, -1, -1, |
| 121 | -1, -1, -1, -1, -1, -1, -1, -1); |
| 122 | } |
| 123 | |
| 124 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 125 | _mm512_zextbf16128_pbh512(__m128bh __a) { |
| 126 | return __builtin_shufflevector( |
| 127 | __a, (__v8bf)_mm_setzero_pbh(), 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, |
| 128 | 13, 14, 15, 8, 9, 10, 11, 12, 13, 14, 15, 8, 9, 10, 11, 12, 13, 14, 15); |
| 129 | } |
| 130 | |
| 131 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 132 | _mm512_zextbf16256_pbh512(__m256bh __a) { |
| 133 | return __builtin_shufflevector(__a, (__v16bf)_mm256_setzero_pbh(), 0, 1, 2, 3, |
| 134 | 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, |
| 135 | 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, |
| 136 | 29, 30, 31); |
| 137 | } |
| 138 | |
| 139 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_abs_pbh(__m512bh __A) { |
| 140 | return (__m512bh)_mm512_and_epi32(_mm512_set1_epi32(0x7FFF7FFF), |
| 141 | (__m512i)__A); |
| 142 | } |
| 143 | |
| 144 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 145 | _mm512_load_pbh(void const *__p) { |
| 146 | return *(const __m512bh *)__p; |
| 147 | } |
| 148 | |
| 149 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 150 | _mm512_loadu_pbh(void const *__p) { |
| 151 | struct __loadu_pbh { |
| 152 | __m512bh_u __v; |
| 153 | } __attribute__((__packed__, __may_alias__)); |
| 154 | return ((const struct __loadu_pbh *)__p)->__v; |
| 155 | } |
| 156 | |
| 157 | static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_store_pbh(void *__P, |
| 158 | __m512bh __A) { |
| 159 | *(__m512bh *)__P = __A; |
| 160 | } |
| 161 | |
| 162 | static __inline__ void __DEFAULT_FN_ATTRS512 _mm512_storeu_pbh(void *__P, |
| 163 | __m512bh __A) { |
| 164 | struct __storeu_pbh { |
| 165 | __m512bh_u __v; |
| 166 | } __attribute__((__packed__, __may_alias__)); |
| 167 | ((struct __storeu_pbh *)__P)->__v = __A; |
| 168 | } |
| 169 | |
| 170 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 171 | _mm512_mask_blend_pbh(__mmask32 __U, __m512bh __A, __m512bh __W) { |
| 172 | return (__m512bh)__builtin_ia32_selectpbf_512((__mmask32)__U, (__v32bf)__W, |
| 173 | (__v32bf)__A); |
| 174 | } |
| 175 | |
| 176 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 177 | _mm512_permutex2var_pbh(__m512bh __A, __m512i __I, __m512bh __B) { |
| 178 | return (__m512bh)__builtin_ia32_vpermi2varhi512((__v32hi)__A, (__v32hi)__I, |
| 179 | (__v32hi)__B); |
| 180 | } |
| 181 | |
| 182 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 183 | _mm512_permutexvar_pbh(__m512i __A, __m512bh __B) { |
| 184 | return (__m512bh)__builtin_ia32_permvarhi512((__v32hi)__B, (__v32hi)__A); |
| 185 | } |
| 186 | |
| 187 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_add_pbh(__m512bh __A, |
| 188 | __m512bh __B) { |
| 189 | return (__m512bh)((__v32bf)__A + (__v32bf)__B); |
| 190 | } |
| 191 | |
| 192 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 193 | _mm512_mask_add_pbh(__m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 194 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 195 | (__mmask32)__U, (__v32bf)_mm512_add_pbh(__A, __B), (__v32bf)__W); |
| 196 | } |
| 197 | |
| 198 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 199 | _mm512_maskz_add_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 200 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 201 | (__mmask32)__U, (__v32bf)_mm512_add_pbh(__A, __B), |
| 202 | (__v32bf)_mm512_setzero_pbh()); |
| 203 | } |
| 204 | |
| 205 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_sub_pbh(__m512bh __A, |
| 206 | __m512bh __B) { |
| 207 | return (__m512bh)((__v32bf)__A - (__v32bf)__B); |
| 208 | } |
| 209 | |
| 210 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 211 | _mm512_mask_sub_pbh(__m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 212 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 213 | (__mmask32)__U, (__v32bf)_mm512_sub_pbh(__A, __B), (__v32bf)__W); |
| 214 | } |
| 215 | |
| 216 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 217 | _mm512_maskz_sub_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 218 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 219 | (__mmask32)__U, (__v32bf)_mm512_sub_pbh(__A, __B), |
| 220 | (__v32bf)_mm512_setzero_pbh()); |
| 221 | } |
| 222 | |
| 223 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mul_pbh(__m512bh __A, |
| 224 | __m512bh __B) { |
| 225 | return (__m512bh)((__v32bf)__A * (__v32bf)__B); |
| 226 | } |
| 227 | |
| 228 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 229 | _mm512_mask_mul_pbh(__m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 230 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 231 | (__mmask32)__U, (__v32bf)_mm512_mul_pbh(__A, __B), (__v32bf)__W); |
| 232 | } |
| 233 | |
| 234 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 235 | _mm512_maskz_mul_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 236 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 237 | (__mmask32)__U, (__v32bf)_mm512_mul_pbh(__A, __B), |
| 238 | (__v32bf)_mm512_setzero_pbh()); |
| 239 | } |
| 240 | |
| 241 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_div_pbh(__m512bh __A, |
| 242 | __m512bh __B) { |
| 243 | return (__m512bh)((__v32bf)__A / (__v32bf)__B); |
| 244 | } |
| 245 | |
| 246 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 247 | _mm512_mask_div_pbh(__m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 248 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 249 | (__mmask32)__U, (__v32bf)_mm512_div_pbh(__A, __B), (__v32bf)__W); |
| 250 | } |
| 251 | |
| 252 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 253 | _mm512_maskz_div_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 254 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 255 | (__mmask32)__U, (__v32bf)_mm512_div_pbh(__A, __B), |
| 256 | (__v32bf)_mm512_setzero_pbh()); |
| 257 | } |
| 258 | |
| 259 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_max_pbh(__m512bh __A, |
| 260 | __m512bh __B) { |
| 261 | return (__m512bh)__builtin_ia32_vmaxbf16512((__v32bf)__A, (__v32bf)__B); |
| 262 | } |
| 263 | |
| 264 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 265 | _mm512_mask_max_pbh(__m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 266 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 267 | (__mmask32)__U, (__v32bf)_mm512_max_pbh(__A, __B), (__v32bf)__W); |
| 268 | } |
| 269 | |
| 270 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 271 | _mm512_maskz_max_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 272 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 273 | (__mmask32)__U, (__v32bf)_mm512_max_pbh(__A, __B), |
| 274 | (__v32bf)_mm512_setzero_pbh()); |
| 275 | } |
| 276 | |
| 277 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_min_pbh(__m512bh __A, |
| 278 | __m512bh __B) { |
| 279 | return (__m512bh)__builtin_ia32_vminbf16512((__v32bf)__A, (__v32bf)__B); |
| 280 | } |
| 281 | |
| 282 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 283 | _mm512_mask_min_pbh(__m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 284 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 285 | (__mmask32)__U, (__v32bf)_mm512_min_pbh(__A, __B), (__v32bf)__W); |
| 286 | } |
| 287 | |
| 288 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 289 | _mm512_maskz_min_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 290 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 291 | (__mmask32)__U, (__v32bf)_mm512_min_pbh(__A, __B), |
| 292 | (__v32bf)_mm512_setzero_pbh()); |
| 293 | } |
| 294 | |
| 295 | #define _mm512_cmp_pbh_mask(__A, __B, __P) \ |
| 296 | ((__mmask32)__builtin_ia32_vcmpbf16512_mask((__v32bf)(__m512bh)(__A), \ |
| 297 | (__v32bf)(__m512bh)(__B), \ |
| 298 | (int)(__P), (__mmask32) - 1)) |
| 299 | |
| 300 | #define _mm512_mask_cmp_pbh_mask(__U, __A, __B, __P) \ |
| 301 | ((__mmask32)__builtin_ia32_vcmpbf16512_mask((__v32bf)(__m512bh)(__A), \ |
| 302 | (__v32bf)(__m512bh)(__B), \ |
| 303 | (int)(__P), (__mmask32)(__U))) |
| 304 | |
| 305 | #define _mm512_mask_fpclass_pbh_mask(__U, __A, imm) \ |
| 306 | ((__mmask32)__builtin_ia32_vfpclassbf16512_mask( \ |
| 307 | (__v32bf)(__m512bh)(__A), (int)(imm), (__mmask32)(__U))) |
| 308 | |
| 309 | #define _mm512_fpclass_pbh_mask(__A, imm) \ |
| 310 | ((__mmask32)__builtin_ia32_vfpclassbf16512_mask( \ |
| 311 | (__v32bf)(__m512bh)(__A), (int)(imm), (__mmask32) - 1)) |
| 312 | |
| 313 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 314 | _mm512_scalef_pbh(__m512bh __A, __m512bh __B) { |
| 315 | return (__m512bh)__builtin_ia32_vscalefbf16512_mask( |
| 316 | (__v32bf)__A, (__v32bf)__B, (__v32bf)_mm512_undefined_pbh(), |
| 317 | (__mmask32)-1); |
| 318 | } |
| 319 | |
| 320 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask_scalef_pbh( |
| 321 | __m512bh __W, __mmask32 __U, __m512bh __A, __m512bh __B) { |
| 322 | return (__m512bh)__builtin_ia32_vscalefbf16512_mask( |
| 323 | (__v32bf)__A, (__v32bf)__B, (__v32bf)__W, (__mmask32)__U); |
| 324 | } |
| 325 | |
| 326 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 327 | _mm512_maskz_scalef_pbh(__mmask32 __U, __m512bh __A, __m512bh __B) { |
| 328 | return (__m512bh)__builtin_ia32_vscalefbf16512_mask( |
| 329 | (__v32bf)__A, (__v32bf)__B, (__v32bf)_mm512_setzero_pbh(), |
| 330 | (__mmask32)__U); |
| 331 | } |
| 332 | |
| 333 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_rcp_pbh(__m512bh __A) { |
| 334 | return (__m512bh)__builtin_ia32_vrcpbf16512_mask( |
| 335 | (__v32bf)__A, (__v32bf)_mm512_undefined_pbh(), (__mmask32)-1); |
| 336 | } |
| 337 | |
| 338 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 339 | _mm512_mask_rcp_pbh(__m512bh __W, __mmask32 __U, __m512bh __A) { |
| 340 | return (__m512bh)__builtin_ia32_vrcpbf16512_mask((__v32bf)__A, (__v32bf)__W, |
| 341 | (__mmask32)__U); |
| 342 | } |
| 343 | |
| 344 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 345 | _mm512_maskz_rcp_pbh(__mmask32 __U, __m512bh __A) { |
| 346 | return (__m512bh)__builtin_ia32_vrcpbf16512_mask( |
| 347 | (__v32bf)__A, (__v32bf)_mm512_setzero_pbh(), (__mmask32)__U); |
| 348 | } |
| 349 | |
| 350 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 351 | _mm512_getexp_pbh(__m512bh __A) { |
| 352 | return (__m512bh)__builtin_ia32_vgetexpbf16512_mask( |
| 353 | (__v32bf)__A, (__v32bf)_mm512_undefined_pbh(), (__mmask32)-1); |
| 354 | } |
| 355 | |
| 356 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 357 | _mm512_mask_getexp_pbh(__m512bh __W, __mmask32 __U, __m512bh __A) { |
| 358 | return (__m512bh)__builtin_ia32_vgetexpbf16512_mask( |
| 359 | (__v32bf)__A, (__v32bf)__W, (__mmask32)__U); |
| 360 | } |
| 361 | |
| 362 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 363 | _mm512_maskz_getexp_pbh(__mmask32 __U, __m512bh __A) { |
| 364 | return (__m512bh)__builtin_ia32_vgetexpbf16512_mask( |
| 365 | (__v32bf)__A, (__v32bf)_mm512_setzero_pbh(), (__mmask32)__U); |
| 366 | } |
| 367 | |
| 368 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 369 | _mm512_rsqrt_pbh(__m512bh __A) { |
| 370 | return (__m512bh)__builtin_ia32_vrsqrtbf16512_mask( |
| 371 | (__v32bf)__A, (__v32bf)_mm512_undefined_pbh(), (__mmask32)-1); |
| 372 | } |
| 373 | |
| 374 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 375 | _mm512_mask_rsqrt_pbh(__m512bh __W, __mmask32 __U, __m512bh __A) { |
| 376 | return (__m512bh)__builtin_ia32_vrsqrtbf16512_mask((__v32bf)__A, (__v32bf)__W, |
| 377 | (__mmask32)__U); |
| 378 | } |
| 379 | |
| 380 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 381 | _mm512_maskz_rsqrt_pbh(__mmask32 __U, __m512bh __A) { |
| 382 | return (__m512bh)__builtin_ia32_vrsqrtbf16512_mask( |
| 383 | (__v32bf)__A, (__v32bf)_mm512_setzero_pbh(), (__mmask32)__U); |
| 384 | } |
| 385 | |
| 386 | #define _mm512_reduce_pbh(__A, imm) \ |
| 387 | ((__m512bh)__builtin_ia32_vreducebf16512_mask( \ |
| 388 | (__v32bf)(__m512bh)(__A), (int)(imm), (__v32bf)_mm512_undefined_pbh(), \ |
| 389 | (__mmask32) - 1)) |
| 390 | |
| 391 | #define _mm512_mask_reduce_pbh(__W, __U, __A, imm) \ |
| 392 | ((__m512bh)__builtin_ia32_vreducebf16512_mask( \ |
| 393 | (__v32bf)(__m512bh)(__A), (int)(imm), (__v32bf)(__m512bh)(__W), \ |
| 394 | (__mmask32)(__U))) |
| 395 | |
| 396 | #define _mm512_maskz_reduce_pbh(__U, __A, imm) \ |
| 397 | ((__m512bh)__builtin_ia32_vreducebf16512_mask( \ |
| 398 | (__v32bf)(__m512bh)(__A), (int)(imm), (__v32bf)_mm512_setzero_pbh(), \ |
| 399 | (__mmask32)(__U))) |
| 400 | |
| 401 | #define _mm512_roundscale_pbh(__A, imm) \ |
| 402 | ((__m512bh)__builtin_ia32_vrndscalebf16_mask( \ |
| 403 | (__v32bf)(__m512bh)(__A), (int)(imm), (__v32bf)_mm512_setzero_pbh(), \ |
| 404 | (__mmask32) - 1)) |
| 405 | |
| 406 | #define _mm512_mask_roundscale_pbh(__W, __U, __A, imm) \ |
| 407 | ((__m512bh)__builtin_ia32_vrndscalebf16_mask( \ |
| 408 | (__v32bf)(__m512bh)(__A), (int)(imm), (__v32bf)(__m512bh)(__W), \ |
| 409 | (__mmask32)(__U))) |
| 410 | |
| 411 | #define _mm512_maskz_roundscale_pbh(__U, __A, imm) \ |
| 412 | ((__m512bh)__builtin_ia32_vrndscalebf16_mask( \ |
| 413 | (__v32bf)(__m512bh)(__A), (int)(imm), (__v32bf)_mm512_setzero_pbh(), \ |
| 414 | (__mmask32)(__U))) |
| 415 | |
| 416 | #define _mm512_getmant_pbh(__A, __B, __C) \ |
| 417 | ((__m512bh)__builtin_ia32_vgetmantbf16512_mask( \ |
| 418 | (__v32bf)(__m512bh)(__A), (int)(((__C) << 2) | (__B)), \ |
| 419 | (__v32bf)_mm512_undefined_pbh(), (__mmask32) - 1)) |
| 420 | |
| 421 | #define _mm512_mask_getmant_pbh(__W, __U, __A, __B, __C) \ |
| 422 | ((__m512bh)__builtin_ia32_vgetmantbf16512_mask( \ |
| 423 | (__v32bf)(__m512bh)(__A), (int)(((__C) << 2) | (__B)), \ |
| 424 | (__v32bf)(__m512bh)(__W), (__mmask32)(__U))) |
| 425 | |
| 426 | #define _mm512_maskz_getmant_pbh(__U, __A, __B, __C) \ |
| 427 | ((__m512bh)__builtin_ia32_vgetmantbf16512_mask( \ |
| 428 | (__v32bf)(__m512bh)(__A), (int)(((__C) << 2) | (__B)), \ |
| 429 | (__v32bf)_mm512_setzero_pbh(), (__mmask32)(__U))) |
| 430 | |
| 431 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_sqrt_pbh(__m512bh __A) { |
| 432 | return __builtin_elementwise_sqrt(__A); |
| 433 | } |
| 434 | |
| 435 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 436 | _mm512_mask_sqrt_pbh(__m512bh __W, __mmask32 __U, __m512bh __A) { |
| 437 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 438 | (__mmask32)__U, (__v32bf)_mm512_sqrt_pbh(__A), (__v32bf)__W); |
| 439 | } |
| 440 | |
| 441 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 442 | _mm512_maskz_sqrt_pbh(__mmask32 __U, __m512bh __A) { |
| 443 | return (__m512bh)__builtin_ia32_selectpbf_512((__mmask32)__U, |
| 444 | (__v32bf)_mm512_sqrt_pbh(__A), |
| 445 | (__v32bf)_mm512_setzero_pbh()); |
| 446 | } |
| 447 | |
| 448 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 449 | _mm512_fmadd_pbh(__m512bh __A, __m512bh __B, __m512bh __C) { |
| 450 | return (__m512bh)__builtin_elementwise_fma((__v32bf)__A, (__v32bf)__B, |
| 451 | (__v32bf)__C); |
| 452 | } |
| 453 | |
| 454 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 455 | _mm512_mask_fmadd_pbh(__m512bh __A, __mmask32 __U, __m512bh __B, __m512bh __C) { |
| 456 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 457 | (__mmask32)__U, |
| 458 | _mm512_fmadd_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), (__v32bf)__A); |
| 459 | } |
| 460 | |
| 461 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask3_fmadd_pbh( |
| 462 | __m512bh __A, __m512bh __B, __m512bh __C, __mmask32 __U) { |
| 463 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 464 | (__mmask32)__U, |
| 465 | _mm512_fmadd_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), (__v32bf)__C); |
| 466 | } |
| 467 | |
| 468 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_fmadd_pbh( |
| 469 | __mmask32 __U, __m512bh __A, __m512bh __B, __m512bh __C) { |
| 470 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 471 | (__mmask32)__U, |
| 472 | _mm512_fmadd_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 473 | (__v32bf)_mm512_setzero_pbh()); |
| 474 | } |
| 475 | |
| 476 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 477 | _mm512_fmsub_pbh(__m512bh __A, __m512bh __B, __m512bh __C) { |
| 478 | return (__m512bh)__builtin_elementwise_fma((__v32bf)__A, (__v32bf)__B, |
| 479 | -(__v32bf)__C); |
| 480 | } |
| 481 | |
| 482 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 483 | _mm512_mask_fmsub_pbh(__m512bh __A, __mmask32 __U, __m512bh __B, __m512bh __C) { |
| 484 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 485 | (__mmask32)__U, |
| 486 | _mm512_fmsub_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), (__v32bf)__A); |
| 487 | } |
| 488 | |
| 489 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask3_fmsub_pbh( |
| 490 | __m512bh __A, __m512bh __B, __m512bh __C, __mmask32 __U) { |
| 491 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 492 | (__mmask32)__U, |
| 493 | _mm512_fmsub_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), (__v32bf)__C); |
| 494 | } |
| 495 | |
| 496 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_fmsub_pbh( |
| 497 | __mmask32 __U, __m512bh __A, __m512bh __B, __m512bh __C) { |
| 498 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 499 | (__mmask32)__U, |
| 500 | _mm512_fmsub_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 501 | (__v32bf)_mm512_setzero_pbh()); |
| 502 | } |
| 503 | |
| 504 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 505 | _mm512_fnmadd_pbh(__m512bh __A, __m512bh __B, __m512bh __C) { |
| 506 | return (__m512bh)__builtin_elementwise_fma((__v32bf)__A, -(__v32bf)__B, |
| 507 | (__v32bf)__C); |
| 508 | } |
| 509 | |
| 510 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask_fnmadd_pbh( |
| 511 | __m512bh __A, __mmask32 __U, __m512bh __B, __m512bh __C) { |
| 512 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 513 | (__mmask32)__U, |
| 514 | _mm512_fnmadd_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 515 | (__v32bf)__A); |
| 516 | } |
| 517 | |
| 518 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask3_fnmadd_pbh( |
| 519 | __m512bh __A, __m512bh __B, __m512bh __C, __mmask32 __U) { |
| 520 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 521 | (__mmask32)__U, |
| 522 | _mm512_fnmadd_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 523 | (__v32bf)__C); |
| 524 | } |
| 525 | |
| 526 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_fnmadd_pbh( |
| 527 | __mmask32 __U, __m512bh __A, __m512bh __B, __m512bh __C) { |
| 528 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 529 | (__mmask32)__U, |
| 530 | _mm512_fnmadd_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 531 | (__v32bf)_mm512_setzero_pbh()); |
| 532 | } |
| 533 | |
| 534 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 |
| 535 | _mm512_fnmsub_pbh(__m512bh __A, __m512bh __B, __m512bh __C) { |
| 536 | return (__m512bh)__builtin_elementwise_fma((__v32bf)__A, -(__v32bf)__B, |
| 537 | -(__v32bf)__C); |
| 538 | } |
| 539 | |
| 540 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask_fnmsub_pbh( |
| 541 | __m512bh __A, __mmask32 __U, __m512bh __B, __m512bh __C) { |
| 542 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 543 | (__mmask32)__U, |
| 544 | _mm512_fnmsub_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 545 | (__v32bf)__A); |
| 546 | } |
| 547 | |
| 548 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_mask3_fnmsub_pbh( |
| 549 | __m512bh __A, __m512bh __B, __m512bh __C, __mmask32 __U) { |
| 550 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 551 | (__mmask32)__U, |
| 552 | _mm512_fnmsub_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 553 | (__v32bf)__C); |
| 554 | } |
| 555 | |
| 556 | static __inline__ __m512bh __DEFAULT_FN_ATTRS512 _mm512_maskz_fnmsub_pbh( |
| 557 | __mmask32 __U, __m512bh __A, __m512bh __B, __m512bh __C) { |
| 558 | return (__m512bh)__builtin_ia32_selectpbf_512( |
| 559 | (__mmask32)__U, |
| 560 | _mm512_fnmsub_pbh((__v32bf)__A, (__v32bf)__B, (__v32bf)__C), |
| 561 | (__v32bf)_mm512_setzero_pbh()); |
| 562 | } |
| 563 | |
| 564 | #undef __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 565 | #undef __DEFAULT_FN_ATTRS512 |
| 566 | |
| 567 | #endif |
| 568 | #endif |