| 1 | /*===---- avx512dqintrin.h - AVX512DQ 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 | |
| 10 | #ifndef __IMMINTRIN_H |
| 11 | #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif |
| 13 | |
| 14 | #ifndef __AVX512DQINTRIN_H |
| 15 | #define __AVX512DQINTRIN_H |
| 16 | |
| 17 | /* Define the default attributes for the functions in this file. */ |
| 18 | #define __DEFAULT_FN_ATTRS512 \ |
| 19 | __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"), \ |
| 20 | __min_vector_width__(512))) |
| 21 | #define __DEFAULT_FN_ATTRS \ |
| 22 | __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"))) |
| 23 | |
| 24 | #if defined(__cplusplus) && (__cplusplus >= 201103L) |
| 25 | #define __DEFAULT_FN_ATTRS512_CONSTEXPR __DEFAULT_FN_ATTRS512 constexpr |
| 26 | #define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS constexpr |
| 27 | #else |
| 28 | #define __DEFAULT_FN_ATTRS512_CONSTEXPR __DEFAULT_FN_ATTRS512 |
| 29 | #define __DEFAULT_FN_ATTRS_CONSTEXPR __DEFAULT_FN_ATTRS |
| 30 | #endif |
| 31 | |
| 32 | static __inline __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 33 | _knot_mask8(__mmask8 __M) { |
| 34 | return __builtin_ia32_knotqi(__M); |
| 35 | } |
| 36 | |
| 37 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 38 | _kand_mask8(__mmask8 __A, __mmask8 __B) { |
| 39 | return (__mmask8)__builtin_ia32_kandqi((__mmask8)__A, (__mmask8)__B); |
| 40 | } |
| 41 | |
| 42 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 43 | _kandn_mask8(__mmask8 __A, __mmask8 __B) { |
| 44 | return (__mmask8)__builtin_ia32_kandnqi((__mmask8)__A, (__mmask8)__B); |
| 45 | } |
| 46 | |
| 47 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 48 | _kor_mask8(__mmask8 __A, __mmask8 __B) { |
| 49 | return (__mmask8)__builtin_ia32_korqi((__mmask8)__A, (__mmask8)__B); |
| 50 | } |
| 51 | |
| 52 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 53 | _kxnor_mask8(__mmask8 __A, __mmask8 __B) { |
| 54 | return (__mmask8)__builtin_ia32_kxnorqi((__mmask8)__A, (__mmask8)__B); |
| 55 | } |
| 56 | |
| 57 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 58 | _kxor_mask8(__mmask8 __A, __mmask8 __B) { |
| 59 | return (__mmask8)__builtin_ia32_kxorqi((__mmask8)__A, (__mmask8)__B); |
| 60 | } |
| 61 | |
| 62 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 63 | _kortestc_mask8_u8(__mmask8 __A, __mmask8 __B) { |
| 64 | return (unsigned char)__builtin_ia32_kortestcqi(__A, __B); |
| 65 | } |
| 66 | |
| 67 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 68 | _kortestz_mask8_u8(__mmask8 __A, __mmask8 __B) { |
| 69 | return (unsigned char)__builtin_ia32_kortestzqi(__A, __B); |
| 70 | } |
| 71 | |
| 72 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 73 | _kortest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) { |
| 74 | *__C = (unsigned char)__builtin_ia32_kortestcqi(__A, __B); |
| 75 | return (unsigned char)__builtin_ia32_kortestzqi(__A, __B); |
| 76 | } |
| 77 | |
| 78 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 79 | _ktestc_mask8_u8(__mmask8 __A, __mmask8 __B) { |
| 80 | return (unsigned char)__builtin_ia32_ktestcqi(__A, __B); |
| 81 | } |
| 82 | |
| 83 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 84 | _ktestz_mask8_u8(__mmask8 __A, __mmask8 __B) { |
| 85 | return (unsigned char)__builtin_ia32_ktestzqi(__A, __B); |
| 86 | } |
| 87 | |
| 88 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 89 | _ktest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) { |
| 90 | *__C = (unsigned char)__builtin_ia32_ktestcqi(__A, __B); |
| 91 | return (unsigned char)__builtin_ia32_ktestzqi(__A, __B); |
| 92 | } |
| 93 | |
| 94 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 95 | _ktestc_mask16_u8(__mmask16 __A, __mmask16 __B) { |
| 96 | return (unsigned char)__builtin_ia32_ktestchi(__A, __B); |
| 97 | } |
| 98 | |
| 99 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 100 | _ktestz_mask16_u8(__mmask16 __A, __mmask16 __B) { |
| 101 | return (unsigned char)__builtin_ia32_ktestzhi(__A, __B); |
| 102 | } |
| 103 | |
| 104 | static __inline__ unsigned char __DEFAULT_FN_ATTRS_CONSTEXPR |
| 105 | _ktest_mask16_u8(__mmask16 __A, __mmask16 __B, unsigned char *__C) { |
| 106 | *__C = (unsigned char)__builtin_ia32_ktestchi(__A, __B); |
| 107 | return (unsigned char)__builtin_ia32_ktestzhi(__A, __B); |
| 108 | } |
| 109 | |
| 110 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 111 | _kadd_mask8(__mmask8 __A, __mmask8 __B) { |
| 112 | return (__mmask8)__builtin_ia32_kaddqi((__mmask8)__A, (__mmask8)__B); |
| 113 | } |
| 114 | |
| 115 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 116 | _kadd_mask16(__mmask16 __A, __mmask16 __B) { |
| 117 | return (__mmask16)__builtin_ia32_kaddhi((__mmask16)__A, (__mmask16)__B); |
| 118 | } |
| 119 | |
| 120 | #define _kshiftli_mask8(A, I) \ |
| 121 | ((__mmask8)__builtin_ia32_kshiftliqi((__mmask8)(A), (unsigned int)(I))) |
| 122 | |
| 123 | #define _kshiftri_mask8(A, I) \ |
| 124 | ((__mmask8)__builtin_ia32_kshiftriqi((__mmask8)(A), (unsigned int)(I))) |
| 125 | |
| 126 | static __inline__ unsigned int |
| 127 | __DEFAULT_FN_ATTRS_CONSTEXPR _cvtmask8_u32(__mmask8 __A) { |
| 128 | return (unsigned int)__builtin_ia32_kmovb((__mmask8)__A); |
| 129 | } |
| 130 | |
| 131 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS_CONSTEXPR |
| 132 | _cvtu32_mask8(unsigned int __A) { |
| 133 | return (__mmask8)__builtin_ia32_kmovb((__mmask8)__A); |
| 134 | } |
| 135 | |
| 136 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS |
| 137 | _load_mask8(__mmask8 *__A) { |
| 138 | return (__mmask8)__builtin_ia32_kmovb(*(__mmask8 *)__A); |
| 139 | } |
| 140 | |
| 141 | static __inline__ void __DEFAULT_FN_ATTRS |
| 142 | _store_mask8(__mmask8 *__A, __mmask8 __B) { |
| 143 | *(__mmask8 *)__A = __builtin_ia32_kmovb((__mmask8)__B); |
| 144 | } |
| 145 | |
| 146 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 147 | _mm512_mullo_epi64(__m512i __A, __m512i __B) { |
| 148 | return (__m512i) ((__v8du) __A * (__v8du) __B); |
| 149 | } |
| 150 | |
| 151 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 152 | _mm512_mask_mullo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) { |
| 153 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U, |
| 154 | (__v8di)_mm512_mullo_epi64(__A, __B), |
| 155 | (__v8di)__W); |
| 156 | } |
| 157 | |
| 158 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 159 | _mm512_maskz_mullo_epi64(__mmask8 __U, __m512i __A, __m512i __B) { |
| 160 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U, |
| 161 | (__v8di)_mm512_mullo_epi64(__A, __B), |
| 162 | (__v8di)_mm512_setzero_si512()); |
| 163 | } |
| 164 | |
| 165 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 166 | _mm512_xor_pd(__m512d __A, __m512d __B) { |
| 167 | return (__m512d)((__v8du)__A ^ (__v8du)__B); |
| 168 | } |
| 169 | |
| 170 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 171 | _mm512_mask_xor_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 172 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 173 | (__v8df)_mm512_xor_pd(__A, __B), |
| 174 | (__v8df)__W); |
| 175 | } |
| 176 | |
| 177 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 178 | _mm512_maskz_xor_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 179 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 180 | (__v8df)_mm512_xor_pd(__A, __B), |
| 181 | (__v8df)_mm512_setzero_pd()); |
| 182 | } |
| 183 | |
| 184 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 185 | _mm512_xor_ps (__m512 __A, __m512 __B) { |
| 186 | return (__m512)((__v16su)__A ^ (__v16su)__B); |
| 187 | } |
| 188 | |
| 189 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 190 | _mm512_mask_xor_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 191 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 192 | (__v16sf)_mm512_xor_ps(__A, __B), |
| 193 | (__v16sf)__W); |
| 194 | } |
| 195 | |
| 196 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 197 | _mm512_maskz_xor_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 198 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 199 | (__v16sf)_mm512_xor_ps(__A, __B), |
| 200 | (__v16sf)_mm512_setzero_ps()); |
| 201 | } |
| 202 | |
| 203 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 204 | _mm512_or_pd(__m512d __A, __m512d __B) { |
| 205 | return (__m512d)((__v8du)__A | (__v8du)__B); |
| 206 | } |
| 207 | |
| 208 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 209 | _mm512_mask_or_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 210 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 211 | (__v8df)_mm512_or_pd(__A, __B), |
| 212 | (__v8df)__W); |
| 213 | } |
| 214 | |
| 215 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 216 | _mm512_maskz_or_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 217 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 218 | (__v8df)_mm512_or_pd(__A, __B), |
| 219 | (__v8df)_mm512_setzero_pd()); |
| 220 | } |
| 221 | |
| 222 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 223 | _mm512_or_ps(__m512 __A, __m512 __B) { |
| 224 | return (__m512)((__v16su)__A | (__v16su)__B); |
| 225 | } |
| 226 | |
| 227 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 228 | _mm512_mask_or_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 229 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 230 | (__v16sf)_mm512_or_ps(__A, __B), |
| 231 | (__v16sf)__W); |
| 232 | } |
| 233 | |
| 234 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 235 | _mm512_maskz_or_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 236 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 237 | (__v16sf)_mm512_or_ps(__A, __B), |
| 238 | (__v16sf)_mm512_setzero_ps()); |
| 239 | } |
| 240 | |
| 241 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 242 | _mm512_and_pd(__m512d __A, __m512d __B) { |
| 243 | return (__m512d)((__v8du)__A & (__v8du)__B); |
| 244 | } |
| 245 | |
| 246 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 247 | _mm512_mask_and_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 248 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 249 | (__v8df)_mm512_and_pd(__A, __B), |
| 250 | (__v8df)__W); |
| 251 | } |
| 252 | |
| 253 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 254 | _mm512_maskz_and_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 255 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 256 | (__v8df)_mm512_and_pd(__A, __B), |
| 257 | (__v8df)_mm512_setzero_pd()); |
| 258 | } |
| 259 | |
| 260 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 261 | _mm512_and_ps(__m512 __A, __m512 __B) { |
| 262 | return (__m512)((__v16su)__A & (__v16su)__B); |
| 263 | } |
| 264 | |
| 265 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 266 | _mm512_mask_and_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 267 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 268 | (__v16sf)_mm512_and_ps(__A, __B), |
| 269 | (__v16sf)__W); |
| 270 | } |
| 271 | |
| 272 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 273 | _mm512_maskz_and_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 274 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 275 | (__v16sf)_mm512_and_ps(__A, __B), |
| 276 | (__v16sf)_mm512_setzero_ps()); |
| 277 | } |
| 278 | |
| 279 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 280 | _mm512_andnot_pd(__m512d __A, __m512d __B) { |
| 281 | return (__m512d)(~(__v8du)__A & (__v8du)__B); |
| 282 | } |
| 283 | |
| 284 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 285 | _mm512_mask_andnot_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) { |
| 286 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 287 | (__v8df)_mm512_andnot_pd(__A, __B), |
| 288 | (__v8df)__W); |
| 289 | } |
| 290 | |
| 291 | static __inline__ __m512d __DEFAULT_FN_ATTRS512 |
| 292 | _mm512_maskz_andnot_pd(__mmask8 __U, __m512d __A, __m512d __B) { |
| 293 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 294 | (__v8df)_mm512_andnot_pd(__A, __B), |
| 295 | (__v8df)_mm512_setzero_pd()); |
| 296 | } |
| 297 | |
| 298 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 299 | _mm512_andnot_ps(__m512 __A, __m512 __B) { |
| 300 | return (__m512)(~(__v16su)__A & (__v16su)__B); |
| 301 | } |
| 302 | |
| 303 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 304 | _mm512_mask_andnot_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) { |
| 305 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 306 | (__v16sf)_mm512_andnot_ps(__A, __B), |
| 307 | (__v16sf)__W); |
| 308 | } |
| 309 | |
| 310 | static __inline__ __m512 __DEFAULT_FN_ATTRS512 |
| 311 | _mm512_maskz_andnot_ps(__mmask16 __U, __m512 __A, __m512 __B) { |
| 312 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__U, |
| 313 | (__v16sf)_mm512_andnot_ps(__A, __B), |
| 314 | (__v16sf)_mm512_setzero_ps()); |
| 315 | } |
| 316 | |
| 317 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 318 | _mm512_cvtpd_epi64 (__m512d __A) { |
| 319 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
| 320 | (__v8di) _mm512_setzero_si512(), |
| 321 | (__mmask8) -1, |
| 322 | _MM_FROUND_CUR_DIRECTION); |
| 323 | } |
| 324 | |
| 325 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 326 | _mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 327 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
| 328 | (__v8di) __W, |
| 329 | (__mmask8) __U, |
| 330 | _MM_FROUND_CUR_DIRECTION); |
| 331 | } |
| 332 | |
| 333 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 334 | _mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) { |
| 335 | return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A, |
| 336 | (__v8di) _mm512_setzero_si512(), |
| 337 | (__mmask8) __U, |
| 338 | _MM_FROUND_CUR_DIRECTION); |
| 339 | } |
| 340 | |
| 341 | #define _mm512_cvt_roundpd_epi64(A, R) \ |
| 342 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 343 | (__v8di)_mm512_setzero_si512(), \ |
| 344 | (__mmask8)-1, (int)(R))) |
| 345 | |
| 346 | #define _mm512_mask_cvt_roundpd_epi64(W, U, A, R) \ |
| 347 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 348 | (__v8di)(__m512i)(W), \ |
| 349 | (__mmask8)(U), (int)(R))) |
| 350 | |
| 351 | #define _mm512_maskz_cvt_roundpd_epi64(U, A, R) \ |
| 352 | ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 353 | (__v8di)_mm512_setzero_si512(), \ |
| 354 | (__mmask8)(U), (int)(R))) |
| 355 | |
| 356 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 357 | _mm512_cvtpd_epu64 (__m512d __A) { |
| 358 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
| 359 | (__v8di) _mm512_setzero_si512(), |
| 360 | (__mmask8) -1, |
| 361 | _MM_FROUND_CUR_DIRECTION); |
| 362 | } |
| 363 | |
| 364 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 365 | _mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 366 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
| 367 | (__v8di) __W, |
| 368 | (__mmask8) __U, |
| 369 | _MM_FROUND_CUR_DIRECTION); |
| 370 | } |
| 371 | |
| 372 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 373 | _mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) { |
| 374 | return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A, |
| 375 | (__v8di) _mm512_setzero_si512(), |
| 376 | (__mmask8) __U, |
| 377 | _MM_FROUND_CUR_DIRECTION); |
| 378 | } |
| 379 | |
| 380 | #define _mm512_cvt_roundpd_epu64(A, R) \ |
| 381 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 382 | (__v8di)_mm512_setzero_si512(), \ |
| 383 | (__mmask8)-1, (int)(R))) |
| 384 | |
| 385 | #define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) \ |
| 386 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 387 | (__v8di)(__m512i)(W), \ |
| 388 | (__mmask8)(U), (int)(R))) |
| 389 | |
| 390 | #define _mm512_maskz_cvt_roundpd_epu64(U, A, R) \ |
| 391 | ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 392 | (__v8di)_mm512_setzero_si512(), \ |
| 393 | (__mmask8)(U), (int)(R))) |
| 394 | |
| 395 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 396 | _mm512_cvtps_epi64 (__m256 __A) { |
| 397 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
| 398 | (__v8di) _mm512_setzero_si512(), |
| 399 | (__mmask8) -1, |
| 400 | _MM_FROUND_CUR_DIRECTION); |
| 401 | } |
| 402 | |
| 403 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 404 | _mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 405 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
| 406 | (__v8di) __W, |
| 407 | (__mmask8) __U, |
| 408 | _MM_FROUND_CUR_DIRECTION); |
| 409 | } |
| 410 | |
| 411 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 412 | _mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) { |
| 413 | return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A, |
| 414 | (__v8di) _mm512_setzero_si512(), |
| 415 | (__mmask8) __U, |
| 416 | _MM_FROUND_CUR_DIRECTION); |
| 417 | } |
| 418 | |
| 419 | #define _mm512_cvt_roundps_epi64(A, R) \ |
| 420 | ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ |
| 421 | (__v8di)_mm512_setzero_si512(), \ |
| 422 | (__mmask8)-1, (int)(R))) |
| 423 | |
| 424 | #define _mm512_mask_cvt_roundps_epi64(W, U, A, R) \ |
| 425 | ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ |
| 426 | (__v8di)(__m512i)(W), \ |
| 427 | (__mmask8)(U), (int)(R))) |
| 428 | |
| 429 | #define _mm512_maskz_cvt_roundps_epi64(U, A, R) \ |
| 430 | ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ |
| 431 | (__v8di)_mm512_setzero_si512(), \ |
| 432 | (__mmask8)(U), (int)(R))) |
| 433 | |
| 434 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 435 | _mm512_cvtps_epu64 (__m256 __A) { |
| 436 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
| 437 | (__v8di) _mm512_setzero_si512(), |
| 438 | (__mmask8) -1, |
| 439 | _MM_FROUND_CUR_DIRECTION); |
| 440 | } |
| 441 | |
| 442 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 443 | _mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 444 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
| 445 | (__v8di) __W, |
| 446 | (__mmask8) __U, |
| 447 | _MM_FROUND_CUR_DIRECTION); |
| 448 | } |
| 449 | |
| 450 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 451 | _mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) { |
| 452 | return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A, |
| 453 | (__v8di) _mm512_setzero_si512(), |
| 454 | (__mmask8) __U, |
| 455 | _MM_FROUND_CUR_DIRECTION); |
| 456 | } |
| 457 | |
| 458 | #define _mm512_cvt_roundps_epu64(A, R) \ |
| 459 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 460 | (__v8di)_mm512_setzero_si512(), \ |
| 461 | (__mmask8)-1, (int)(R))) |
| 462 | |
| 463 | #define _mm512_mask_cvt_roundps_epu64(W, U, A, R) \ |
| 464 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 465 | (__v8di)(__m512i)(W), \ |
| 466 | (__mmask8)(U), (int)(R))) |
| 467 | |
| 468 | #define _mm512_maskz_cvt_roundps_epu64(U, A, R) \ |
| 469 | ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 470 | (__v8di)_mm512_setzero_si512(), \ |
| 471 | (__mmask8)(U), (int)(R))) |
| 472 | |
| 473 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 474 | _mm512_cvtepi64_pd(__m512i __A) { |
| 475 | return (__m512d)__builtin_convertvector((__v8di)__A, __v8df); |
| 476 | } |
| 477 | |
| 478 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 479 | _mm512_mask_cvtepi64_pd(__m512d __W, __mmask8 __U, __m512i __A) { |
| 480 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 481 | (__v8df)_mm512_cvtepi64_pd(__A), |
| 482 | (__v8df)__W); |
| 483 | } |
| 484 | |
| 485 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 486 | _mm512_maskz_cvtepi64_pd(__mmask8 __U, __m512i __A) { |
| 487 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 488 | (__v8df)_mm512_cvtepi64_pd(__A), |
| 489 | (__v8df)_mm512_setzero_pd()); |
| 490 | } |
| 491 | |
| 492 | #define _mm512_cvt_roundepi64_pd(A, R) \ |
| 493 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 494 | (__v8df)_mm512_setzero_pd(), \ |
| 495 | (__mmask8)-1, (int)(R))) |
| 496 | |
| 497 | #define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) \ |
| 498 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 499 | (__v8df)(__m512d)(W), \ |
| 500 | (__mmask8)(U), (int)(R))) |
| 501 | |
| 502 | #define _mm512_maskz_cvt_roundepi64_pd(U, A, R) \ |
| 503 | ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 504 | (__v8df)_mm512_setzero_pd(), \ |
| 505 | (__mmask8)(U), (int)(R))) |
| 506 | |
| 507 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
| 508 | _mm512_cvtepi64_ps (__m512i __A) { |
| 509 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
| 510 | (__v8sf) _mm256_setzero_ps(), |
| 511 | (__mmask8) -1, |
| 512 | _MM_FROUND_CUR_DIRECTION); |
| 513 | } |
| 514 | |
| 515 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
| 516 | _mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) { |
| 517 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
| 518 | (__v8sf) __W, |
| 519 | (__mmask8) __U, |
| 520 | _MM_FROUND_CUR_DIRECTION); |
| 521 | } |
| 522 | |
| 523 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
| 524 | _mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) { |
| 525 | return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A, |
| 526 | (__v8sf) _mm256_setzero_ps(), |
| 527 | (__mmask8) __U, |
| 528 | _MM_FROUND_CUR_DIRECTION); |
| 529 | } |
| 530 | |
| 531 | #define _mm512_cvt_roundepi64_ps(A, R) \ |
| 532 | ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 533 | (__v8sf)_mm256_setzero_ps(), \ |
| 534 | (__mmask8)-1, (int)(R))) |
| 535 | |
| 536 | #define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) \ |
| 537 | ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 538 | (__v8sf)(__m256)(W), (__mmask8)(U), \ |
| 539 | (int)(R))) |
| 540 | |
| 541 | #define _mm512_maskz_cvt_roundepi64_ps(U, A, R) \ |
| 542 | ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 543 | (__v8sf)_mm256_setzero_ps(), \ |
| 544 | (__mmask8)(U), (int)(R))) |
| 545 | |
| 546 | |
| 547 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 548 | _mm512_cvttpd_epi64 (__m512d __A) { |
| 549 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
| 550 | (__v8di) _mm512_setzero_si512(), |
| 551 | (__mmask8) -1, |
| 552 | _MM_FROUND_CUR_DIRECTION); |
| 553 | } |
| 554 | |
| 555 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 556 | _mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 557 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
| 558 | (__v8di) __W, |
| 559 | (__mmask8) __U, |
| 560 | _MM_FROUND_CUR_DIRECTION); |
| 561 | } |
| 562 | |
| 563 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 564 | _mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) { |
| 565 | return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A, |
| 566 | (__v8di) _mm512_setzero_si512(), |
| 567 | (__mmask8) __U, |
| 568 | _MM_FROUND_CUR_DIRECTION); |
| 569 | } |
| 570 | |
| 571 | #define _mm512_cvtt_roundpd_epi64(A, R) \ |
| 572 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 573 | (__v8di)_mm512_setzero_si512(), \ |
| 574 | (__mmask8)-1, (int)(R))) |
| 575 | |
| 576 | #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) \ |
| 577 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 578 | (__v8di)(__m512i)(W), \ |
| 579 | (__mmask8)(U), (int)(R))) |
| 580 | |
| 581 | #define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) \ |
| 582 | ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ |
| 583 | (__v8di)_mm512_setzero_si512(), \ |
| 584 | (__mmask8)(U), (int)(R))) |
| 585 | |
| 586 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 587 | _mm512_cvttpd_epu64 (__m512d __A) { |
| 588 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
| 589 | (__v8di) _mm512_setzero_si512(), |
| 590 | (__mmask8) -1, |
| 591 | _MM_FROUND_CUR_DIRECTION); |
| 592 | } |
| 593 | |
| 594 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 595 | _mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) { |
| 596 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
| 597 | (__v8di) __W, |
| 598 | (__mmask8) __U, |
| 599 | _MM_FROUND_CUR_DIRECTION); |
| 600 | } |
| 601 | |
| 602 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 603 | _mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) { |
| 604 | return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A, |
| 605 | (__v8di) _mm512_setzero_si512(), |
| 606 | (__mmask8) __U, |
| 607 | _MM_FROUND_CUR_DIRECTION); |
| 608 | } |
| 609 | |
| 610 | #define _mm512_cvtt_roundpd_epu64(A, R) \ |
| 611 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 612 | (__v8di)_mm512_setzero_si512(), \ |
| 613 | (__mmask8)-1, (int)(R))) |
| 614 | |
| 615 | #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) \ |
| 616 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 617 | (__v8di)(__m512i)(W), \ |
| 618 | (__mmask8)(U), (int)(R))) |
| 619 | |
| 620 | #define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) \ |
| 621 | ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ |
| 622 | (__v8di)_mm512_setzero_si512(), \ |
| 623 | (__mmask8)(U), (int)(R))) |
| 624 | |
| 625 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 626 | _mm512_cvttps_epi64 (__m256 __A) { |
| 627 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
| 628 | (__v8di) _mm512_setzero_si512(), |
| 629 | (__mmask8) -1, |
| 630 | _MM_FROUND_CUR_DIRECTION); |
| 631 | } |
| 632 | |
| 633 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 634 | _mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 635 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
| 636 | (__v8di) __W, |
| 637 | (__mmask8) __U, |
| 638 | _MM_FROUND_CUR_DIRECTION); |
| 639 | } |
| 640 | |
| 641 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 642 | _mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) { |
| 643 | return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A, |
| 644 | (__v8di) _mm512_setzero_si512(), |
| 645 | (__mmask8) __U, |
| 646 | _MM_FROUND_CUR_DIRECTION); |
| 647 | } |
| 648 | |
| 649 | #define _mm512_cvtt_roundps_epi64(A, R) \ |
| 650 | ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ |
| 651 | (__v8di)_mm512_setzero_si512(), \ |
| 652 | (__mmask8)-1, (int)(R))) |
| 653 | |
| 654 | #define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) \ |
| 655 | ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ |
| 656 | (__v8di)(__m512i)(W), \ |
| 657 | (__mmask8)(U), (int)(R))) |
| 658 | |
| 659 | #define _mm512_maskz_cvtt_roundps_epi64(U, A, R) \ |
| 660 | ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ |
| 661 | (__v8di)_mm512_setzero_si512(), \ |
| 662 | (__mmask8)(U), (int)(R))) |
| 663 | |
| 664 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 665 | _mm512_cvttps_epu64 (__m256 __A) { |
| 666 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
| 667 | (__v8di) _mm512_setzero_si512(), |
| 668 | (__mmask8) -1, |
| 669 | _MM_FROUND_CUR_DIRECTION); |
| 670 | } |
| 671 | |
| 672 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 673 | _mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) { |
| 674 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
| 675 | (__v8di) __W, |
| 676 | (__mmask8) __U, |
| 677 | _MM_FROUND_CUR_DIRECTION); |
| 678 | } |
| 679 | |
| 680 | static __inline__ __m512i __DEFAULT_FN_ATTRS512 |
| 681 | _mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) { |
| 682 | return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A, |
| 683 | (__v8di) _mm512_setzero_si512(), |
| 684 | (__mmask8) __U, |
| 685 | _MM_FROUND_CUR_DIRECTION); |
| 686 | } |
| 687 | |
| 688 | #define _mm512_cvtt_roundps_epu64(A, R) \ |
| 689 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 690 | (__v8di)_mm512_setzero_si512(), \ |
| 691 | (__mmask8)-1, (int)(R))) |
| 692 | |
| 693 | #define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) \ |
| 694 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 695 | (__v8di)(__m512i)(W), \ |
| 696 | (__mmask8)(U), (int)(R))) |
| 697 | |
| 698 | #define _mm512_maskz_cvtt_roundps_epu64(U, A, R) \ |
| 699 | ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ |
| 700 | (__v8di)_mm512_setzero_si512(), \ |
| 701 | (__mmask8)(U), (int)(R))) |
| 702 | |
| 703 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 704 | _mm512_cvtepu64_pd(__m512i __A) { |
| 705 | return (__m512d)__builtin_convertvector((__v8du)__A, __v8df); |
| 706 | } |
| 707 | |
| 708 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 709 | _mm512_mask_cvtepu64_pd(__m512d __W, __mmask8 __U, __m512i __A) { |
| 710 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 711 | (__v8df)_mm512_cvtepu64_pd(__A), |
| 712 | (__v8df)__W); |
| 713 | } |
| 714 | |
| 715 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 716 | _mm512_maskz_cvtepu64_pd(__mmask8 __U, __m512i __A) { |
| 717 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U, |
| 718 | (__v8df)_mm512_cvtepu64_pd(__A), |
| 719 | (__v8df)_mm512_setzero_pd()); |
| 720 | } |
| 721 | |
| 722 | #define _mm512_cvt_roundepu64_pd(A, R) \ |
| 723 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 724 | (__v8df)_mm512_setzero_pd(), \ |
| 725 | (__mmask8)-1, (int)(R))) |
| 726 | |
| 727 | #define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) \ |
| 728 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 729 | (__v8df)(__m512d)(W), \ |
| 730 | (__mmask8)(U), (int)(R))) |
| 731 | |
| 732 | |
| 733 | #define _mm512_maskz_cvt_roundepu64_pd(U, A, R) \ |
| 734 | ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ |
| 735 | (__v8df)_mm512_setzero_pd(), \ |
| 736 | (__mmask8)(U), (int)(R))) |
| 737 | |
| 738 | |
| 739 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
| 740 | _mm512_cvtepu64_ps (__m512i __A) { |
| 741 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
| 742 | (__v8sf) _mm256_setzero_ps(), |
| 743 | (__mmask8) -1, |
| 744 | _MM_FROUND_CUR_DIRECTION); |
| 745 | } |
| 746 | |
| 747 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
| 748 | _mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) { |
| 749 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
| 750 | (__v8sf) __W, |
| 751 | (__mmask8) __U, |
| 752 | _MM_FROUND_CUR_DIRECTION); |
| 753 | } |
| 754 | |
| 755 | static __inline__ __m256 __DEFAULT_FN_ATTRS512 |
| 756 | _mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) { |
| 757 | return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A, |
| 758 | (__v8sf) _mm256_setzero_ps(), |
| 759 | (__mmask8) __U, |
| 760 | _MM_FROUND_CUR_DIRECTION); |
| 761 | } |
| 762 | |
| 763 | #define _mm512_cvt_roundepu64_ps(A, R) \ |
| 764 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 765 | (__v8sf)_mm256_setzero_ps(), \ |
| 766 | (__mmask8)-1, (int)(R))) |
| 767 | |
| 768 | #define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) \ |
| 769 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 770 | (__v8sf)(__m256)(W), (__mmask8)(U), \ |
| 771 | (int)(R))) |
| 772 | |
| 773 | #define _mm512_maskz_cvt_roundepu64_ps(U, A, R) \ |
| 774 | ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ |
| 775 | (__v8sf)_mm256_setzero_ps(), \ |
| 776 | (__mmask8)(U), (int)(R))) |
| 777 | |
| 778 | #define _mm512_range_pd(A, B, C) \ |
| 779 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 780 | (__v8df)(__m512d)(B), (int)(C), \ |
| 781 | (__v8df)_mm512_setzero_pd(), \ |
| 782 | (__mmask8)-1, \ |
| 783 | _MM_FROUND_CUR_DIRECTION)) |
| 784 | |
| 785 | #define _mm512_mask_range_pd(W, U, A, B, C) \ |
| 786 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 787 | (__v8df)(__m512d)(B), (int)(C), \ |
| 788 | (__v8df)(__m512d)(W), (__mmask8)(U), \ |
| 789 | _MM_FROUND_CUR_DIRECTION)) |
| 790 | |
| 791 | #define _mm512_maskz_range_pd(U, A, B, C) \ |
| 792 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 793 | (__v8df)(__m512d)(B), (int)(C), \ |
| 794 | (__v8df)_mm512_setzero_pd(), \ |
| 795 | (__mmask8)(U), \ |
| 796 | _MM_FROUND_CUR_DIRECTION)) |
| 797 | |
| 798 | #define _mm512_range_round_pd(A, B, C, R) \ |
| 799 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 800 | (__v8df)(__m512d)(B), (int)(C), \ |
| 801 | (__v8df)_mm512_setzero_pd(), \ |
| 802 | (__mmask8)-1, (int)(R))) |
| 803 | |
| 804 | #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ |
| 805 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 806 | (__v8df)(__m512d)(B), (int)(C), \ |
| 807 | (__v8df)(__m512d)(W), (__mmask8)(U), \ |
| 808 | (int)(R))) |
| 809 | |
| 810 | #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ |
| 811 | ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ |
| 812 | (__v8df)(__m512d)(B), (int)(C), \ |
| 813 | (__v8df)_mm512_setzero_pd(), \ |
| 814 | (__mmask8)(U), (int)(R))) |
| 815 | |
| 816 | #define _mm512_range_ps(A, B, C) \ |
| 817 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 818 | (__v16sf)(__m512)(B), (int)(C), \ |
| 819 | (__v16sf)_mm512_setzero_ps(), \ |
| 820 | (__mmask16)-1, \ |
| 821 | _MM_FROUND_CUR_DIRECTION)) |
| 822 | |
| 823 | #define _mm512_mask_range_ps(W, U, A, B, C) \ |
| 824 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 825 | (__v16sf)(__m512)(B), (int)(C), \ |
| 826 | (__v16sf)(__m512)(W), (__mmask16)(U), \ |
| 827 | _MM_FROUND_CUR_DIRECTION)) |
| 828 | |
| 829 | #define _mm512_maskz_range_ps(U, A, B, C) \ |
| 830 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 831 | (__v16sf)(__m512)(B), (int)(C), \ |
| 832 | (__v16sf)_mm512_setzero_ps(), \ |
| 833 | (__mmask16)(U), \ |
| 834 | _MM_FROUND_CUR_DIRECTION)) |
| 835 | |
| 836 | #define _mm512_range_round_ps(A, B, C, R) \ |
| 837 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 838 | (__v16sf)(__m512)(B), (int)(C), \ |
| 839 | (__v16sf)_mm512_setzero_ps(), \ |
| 840 | (__mmask16)-1, (int)(R))) |
| 841 | |
| 842 | #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ |
| 843 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 844 | (__v16sf)(__m512)(B), (int)(C), \ |
| 845 | (__v16sf)(__m512)(W), (__mmask16)(U), \ |
| 846 | (int)(R))) |
| 847 | |
| 848 | #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ |
| 849 | ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ |
| 850 | (__v16sf)(__m512)(B), (int)(C), \ |
| 851 | (__v16sf)_mm512_setzero_ps(), \ |
| 852 | (__mmask16)(U), (int)(R))) |
| 853 | |
| 854 | #define _mm_range_round_ss(A, B, C, R) \ |
| 855 | ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ |
| 856 | (__v4sf)(__m128)(B), \ |
| 857 | (__v4sf)_mm_setzero_ps(), \ |
| 858 | (__mmask8) -1, (int)(C),\ |
| 859 | (int)(R))) |
| 860 | |
| 861 | #define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 862 | |
| 863 | #define _mm_mask_range_round_ss(W, U, A, B, C, R) \ |
| 864 | ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ |
| 865 | (__v4sf)(__m128)(B), \ |
| 866 | (__v4sf)(__m128)(W),\ |
| 867 | (__mmask8)(U), (int)(C),\ |
| 868 | (int)(R))) |
| 869 | |
| 870 | #define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION) |
| 871 | |
| 872 | #define _mm_maskz_range_round_ss(U, A, B, C, R) \ |
| 873 | ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ |
| 874 | (__v4sf)(__m128)(B), \ |
| 875 | (__v4sf)_mm_setzero_ps(), \ |
| 876 | (__mmask8)(U), (int)(C),\ |
| 877 | (int)(R))) |
| 878 | |
| 879 | #define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 880 | |
| 881 | #define _mm_range_round_sd(A, B, C, R) \ |
| 882 | ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ |
| 883 | (__v2df)(__m128d)(B), \ |
| 884 | (__v2df)_mm_setzero_pd(), \ |
| 885 | (__mmask8) -1, (int)(C),\ |
| 886 | (int)(R))) |
| 887 | |
| 888 | #define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 889 | |
| 890 | #define _mm_mask_range_round_sd(W, U, A, B, C, R) \ |
| 891 | ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ |
| 892 | (__v2df)(__m128d)(B), \ |
| 893 | (__v2df)(__m128d)(W),\ |
| 894 | (__mmask8)(U), (int)(C),\ |
| 895 | (int)(R))) |
| 896 | |
| 897 | #define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 898 | |
| 899 | #define _mm_maskz_range_round_sd(U, A, B, C, R) \ |
| 900 | ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ |
| 901 | (__v2df)(__m128d)(B), \ |
| 902 | (__v2df)_mm_setzero_pd(), \ |
| 903 | (__mmask8)(U), (int)(C),\ |
| 904 | (int)(R))) |
| 905 | |
| 906 | #define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) |
| 907 | |
| 908 | #define _mm512_reduce_pd(A, B) \ |
| 909 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 910 | (__v8df)_mm512_setzero_pd(), \ |
| 911 | (__mmask8)-1, \ |
| 912 | _MM_FROUND_CUR_DIRECTION)) |
| 913 | |
| 914 | #define _mm512_mask_reduce_pd(W, U, A, B) \ |
| 915 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 916 | (__v8df)(__m512d)(W), \ |
| 917 | (__mmask8)(U), \ |
| 918 | _MM_FROUND_CUR_DIRECTION)) |
| 919 | |
| 920 | #define _mm512_maskz_reduce_pd(U, A, B) \ |
| 921 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 922 | (__v8df)_mm512_setzero_pd(), \ |
| 923 | (__mmask8)(U), \ |
| 924 | _MM_FROUND_CUR_DIRECTION)) |
| 925 | |
| 926 | #define _mm512_reduce_ps(A, B) \ |
| 927 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 928 | (__v16sf)_mm512_setzero_ps(), \ |
| 929 | (__mmask16)-1, \ |
| 930 | _MM_FROUND_CUR_DIRECTION)) |
| 931 | |
| 932 | #define _mm512_mask_reduce_ps(W, U, A, B) \ |
| 933 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 934 | (__v16sf)(__m512)(W), \ |
| 935 | (__mmask16)(U), \ |
| 936 | _MM_FROUND_CUR_DIRECTION)) |
| 937 | |
| 938 | #define _mm512_maskz_reduce_ps(U, A, B) \ |
| 939 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 940 | (__v16sf)_mm512_setzero_ps(), \ |
| 941 | (__mmask16)(U), \ |
| 942 | _MM_FROUND_CUR_DIRECTION)) |
| 943 | |
| 944 | #define _mm512_reduce_round_pd(A, B, R) \ |
| 945 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 946 | (__v8df)_mm512_setzero_pd(), \ |
| 947 | (__mmask8)-1, (int)(R))) |
| 948 | |
| 949 | #define _mm512_mask_reduce_round_pd(W, U, A, B, R) \ |
| 950 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 951 | (__v8df)(__m512d)(W), \ |
| 952 | (__mmask8)(U), (int)(R))) |
| 953 | |
| 954 | #define _mm512_maskz_reduce_round_pd(U, A, B, R) \ |
| 955 | ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ |
| 956 | (__v8df)_mm512_setzero_pd(), \ |
| 957 | (__mmask8)(U), (int)(R))) |
| 958 | |
| 959 | #define _mm512_reduce_round_ps(A, B, R) \ |
| 960 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 961 | (__v16sf)_mm512_setzero_ps(), \ |
| 962 | (__mmask16)-1, (int)(R))) |
| 963 | |
| 964 | #define _mm512_mask_reduce_round_ps(W, U, A, B, R) \ |
| 965 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 966 | (__v16sf)(__m512)(W), \ |
| 967 | (__mmask16)(U), (int)(R))) |
| 968 | |
| 969 | #define _mm512_maskz_reduce_round_ps(U, A, B, R) \ |
| 970 | ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ |
| 971 | (__v16sf)_mm512_setzero_ps(), \ |
| 972 | (__mmask16)(U), (int)(R))) |
| 973 | |
| 974 | #define _mm_reduce_ss(A, B, C) \ |
| 975 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 976 | (__v4sf)(__m128)(B), \ |
| 977 | (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ |
| 978 | (int)(C), _MM_FROUND_CUR_DIRECTION)) |
| 979 | |
| 980 | #define _mm_mask_reduce_ss(W, U, A, B, C) \ |
| 981 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 982 | (__v4sf)(__m128)(B), \ |
| 983 | (__v4sf)(__m128)(W), (__mmask8)(U), \ |
| 984 | (int)(C), _MM_FROUND_CUR_DIRECTION)) |
| 985 | |
| 986 | #define _mm_maskz_reduce_ss(U, A, B, C) \ |
| 987 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 988 | (__v4sf)(__m128)(B), \ |
| 989 | (__v4sf)_mm_setzero_ps(), \ |
| 990 | (__mmask8)(U), (int)(C), \ |
| 991 | _MM_FROUND_CUR_DIRECTION)) |
| 992 | |
| 993 | #define _mm_reduce_round_ss(A, B, C, R) \ |
| 994 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 995 | (__v4sf)(__m128)(B), \ |
| 996 | (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ |
| 997 | (int)(C), (int)(R))) |
| 998 | |
| 999 | #define _mm_mask_reduce_round_ss(W, U, A, B, C, R) \ |
| 1000 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 1001 | (__v4sf)(__m128)(B), \ |
| 1002 | (__v4sf)(__m128)(W), (__mmask8)(U), \ |
| 1003 | (int)(C), (int)(R))) |
| 1004 | |
| 1005 | #define _mm_maskz_reduce_round_ss(U, A, B, C, R) \ |
| 1006 | ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ |
| 1007 | (__v4sf)(__m128)(B), \ |
| 1008 | (__v4sf)_mm_setzero_ps(), \ |
| 1009 | (__mmask8)(U), (int)(C), (int)(R))) |
| 1010 | |
| 1011 | #define _mm_reduce_sd(A, B, C) \ |
| 1012 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1013 | (__v2df)(__m128d)(B), \ |
| 1014 | (__v2df)_mm_setzero_pd(), \ |
| 1015 | (__mmask8)-1, (int)(C), \ |
| 1016 | _MM_FROUND_CUR_DIRECTION)) |
| 1017 | |
| 1018 | #define _mm_mask_reduce_sd(W, U, A, B, C) \ |
| 1019 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1020 | (__v2df)(__m128d)(B), \ |
| 1021 | (__v2df)(__m128d)(W), (__mmask8)(U), \ |
| 1022 | (int)(C), _MM_FROUND_CUR_DIRECTION)) |
| 1023 | |
| 1024 | #define _mm_maskz_reduce_sd(U, A, B, C) \ |
| 1025 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1026 | (__v2df)(__m128d)(B), \ |
| 1027 | (__v2df)_mm_setzero_pd(), \ |
| 1028 | (__mmask8)(U), (int)(C), \ |
| 1029 | _MM_FROUND_CUR_DIRECTION)) |
| 1030 | |
| 1031 | #define _mm_reduce_round_sd(A, B, C, R) \ |
| 1032 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1033 | (__v2df)(__m128d)(B), \ |
| 1034 | (__v2df)_mm_setzero_pd(), \ |
| 1035 | (__mmask8)-1, (int)(C), (int)(R))) |
| 1036 | |
| 1037 | #define _mm_mask_reduce_round_sd(W, U, A, B, C, R) \ |
| 1038 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1039 | (__v2df)(__m128d)(B), \ |
| 1040 | (__v2df)(__m128d)(W), (__mmask8)(U), \ |
| 1041 | (int)(C), (int)(R))) |
| 1042 | |
| 1043 | #define _mm_maskz_reduce_round_sd(U, A, B, C, R) \ |
| 1044 | ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ |
| 1045 | (__v2df)(__m128d)(B), \ |
| 1046 | (__v2df)_mm_setzero_pd(), \ |
| 1047 | (__mmask8)(U), (int)(C), (int)(R))) |
| 1048 | |
| 1049 | static __inline__ __mmask16 |
| 1050 | __DEFAULT_FN_ATTRS512_CONSTEXPR _mm512_movepi32_mask(__m512i __A) { |
| 1051 | return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A); |
| 1052 | } |
| 1053 | |
| 1054 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1055 | _mm512_movm_epi32(__mmask16 __A) { |
| 1056 | return (__m512i) __builtin_ia32_cvtmask2d512 (__A); |
| 1057 | } |
| 1058 | |
| 1059 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1060 | _mm512_movm_epi64(__mmask8 __A) { |
| 1061 | return (__m512i) __builtin_ia32_cvtmask2q512 (__A); |
| 1062 | } |
| 1063 | |
| 1064 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1065 | _mm512_movepi64_mask(__m512i __A) { |
| 1066 | return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A); |
| 1067 | } |
| 1068 | |
| 1069 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1070 | _mm512_broadcast_f32x2(__m128 __A) { |
| 1071 | return (__m512)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A, |
| 1072 | 0, 1, 0, 1, 0, 1, 0, 1, |
| 1073 | 0, 1, 0, 1, 0, 1, 0, 1); |
| 1074 | } |
| 1075 | |
| 1076 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1077 | _mm512_mask_broadcast_f32x2(__m512 __O, __mmask16 __M, __m128 __A) { |
| 1078 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1079 | (__v16sf)_mm512_broadcast_f32x2(__A), |
| 1080 | (__v16sf)__O); |
| 1081 | } |
| 1082 | |
| 1083 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1084 | _mm512_maskz_broadcast_f32x2(__mmask16 __M, __m128 __A) { |
| 1085 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1086 | (__v16sf)_mm512_broadcast_f32x2(__A), |
| 1087 | (__v16sf)_mm512_setzero_ps()); |
| 1088 | } |
| 1089 | |
| 1090 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1091 | _mm512_broadcast_f32x8(__m256 __A) { |
| 1092 | return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A, |
| 1093 | 0, 1, 2, 3, 4, 5, 6, 7, |
| 1094 | 0, 1, 2, 3, 4, 5, 6, 7); |
| 1095 | } |
| 1096 | |
| 1097 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1098 | _mm512_mask_broadcast_f32x8(__m512 __O, __mmask16 __M, __m256 __A) { |
| 1099 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1100 | (__v16sf)_mm512_broadcast_f32x8(__A), |
| 1101 | (__v16sf)__O); |
| 1102 | } |
| 1103 | |
| 1104 | static __inline__ __m512 __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1105 | _mm512_maskz_broadcast_f32x8(__mmask16 __M, __m256 __A) { |
| 1106 | return (__m512)__builtin_ia32_selectps_512((__mmask16)__M, |
| 1107 | (__v16sf)_mm512_broadcast_f32x8(__A), |
| 1108 | (__v16sf)_mm512_setzero_ps()); |
| 1109 | } |
| 1110 | |
| 1111 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1112 | _mm512_broadcast_f64x2(__m128d __A) { |
| 1113 | return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A, |
| 1114 | 0, 1, 0, 1, 0, 1, 0, 1); |
| 1115 | } |
| 1116 | |
| 1117 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1118 | _mm512_mask_broadcast_f64x2(__m512d __O, __mmask8 __M, __m128d __A) { |
| 1119 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M, |
| 1120 | (__v8df)_mm512_broadcast_f64x2(__A), |
| 1121 | (__v8df)__O); |
| 1122 | } |
| 1123 | |
| 1124 | static __inline__ __m512d __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1125 | _mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A) { |
| 1126 | return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M, |
| 1127 | (__v8df)_mm512_broadcast_f64x2(__A), |
| 1128 | (__v8df)_mm512_setzero_pd()); |
| 1129 | } |
| 1130 | |
| 1131 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1132 | _mm512_broadcast_i32x2(__m128i __A) { |
| 1133 | return (__m512i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
| 1134 | 0, 1, 0, 1, 0, 1, 0, 1, |
| 1135 | 0, 1, 0, 1, 0, 1, 0, 1); |
| 1136 | } |
| 1137 | |
| 1138 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1139 | _mm512_mask_broadcast_i32x2(__m512i __O, __mmask16 __M, __m128i __A) { |
| 1140 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1141 | (__v16si)_mm512_broadcast_i32x2(__A), |
| 1142 | (__v16si)__O); |
| 1143 | } |
| 1144 | |
| 1145 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1146 | _mm512_maskz_broadcast_i32x2(__mmask16 __M, __m128i __A) { |
| 1147 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1148 | (__v16si)_mm512_broadcast_i32x2(__A), |
| 1149 | (__v16si)_mm512_setzero_si512()); |
| 1150 | } |
| 1151 | |
| 1152 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1153 | _mm512_broadcast_i32x8(__m256i __A) { |
| 1154 | return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A, |
| 1155 | 0, 1, 2, 3, 4, 5, 6, 7, |
| 1156 | 0, 1, 2, 3, 4, 5, 6, 7); |
| 1157 | } |
| 1158 | |
| 1159 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1160 | _mm512_mask_broadcast_i32x8(__m512i __O, __mmask16 __M, __m256i __A) { |
| 1161 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1162 | (__v16si)_mm512_broadcast_i32x8(__A), |
| 1163 | (__v16si)__O); |
| 1164 | } |
| 1165 | |
| 1166 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1167 | _mm512_maskz_broadcast_i32x8(__mmask16 __M, __m256i __A) { |
| 1168 | return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M, |
| 1169 | (__v16si)_mm512_broadcast_i32x8(__A), |
| 1170 | (__v16si)_mm512_setzero_si512()); |
| 1171 | } |
| 1172 | |
| 1173 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1174 | _mm512_broadcast_i64x2(__m128i __A) { |
| 1175 | return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A, |
| 1176 | 0, 1, 0, 1, 0, 1, 0, 1); |
| 1177 | } |
| 1178 | |
| 1179 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1180 | _mm512_mask_broadcast_i64x2(__m512i __O, __mmask8 __M, __m128i __A) { |
| 1181 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, |
| 1182 | (__v8di)_mm512_broadcast_i64x2(__A), |
| 1183 | (__v8di)__O); |
| 1184 | } |
| 1185 | |
| 1186 | static __inline__ __m512i __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1187 | _mm512_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A) { |
| 1188 | return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M, |
| 1189 | (__v8di)_mm512_broadcast_i64x2(__A), |
| 1190 | (__v8di)_mm512_setzero_si512()); |
| 1191 | } |
| 1192 | |
| 1193 | #define _mm512_extractf32x8_ps(A, imm) \ |
| 1194 | ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ |
| 1195 | (__v8sf)_mm256_setzero_ps(), \ |
| 1196 | (__mmask8) - 1)) |
| 1197 | |
| 1198 | #define _mm512_mask_extractf32x8_ps(W, U, A, imm) \ |
| 1199 | ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ |
| 1200 | (__v8sf)(__m256)(W), \ |
| 1201 | (__mmask8)(U))) |
| 1202 | |
| 1203 | #define _mm512_maskz_extractf32x8_ps(U, A, imm) \ |
| 1204 | ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ |
| 1205 | (__v8sf)_mm256_setzero_ps(), \ |
| 1206 | (__mmask8)(U))) |
| 1207 | |
| 1208 | #define _mm512_extractf64x2_pd(A, imm) \ |
| 1209 | ((__m128d)__builtin_ia32_extractf64x2_512_mask( \ |
| 1210 | (__v8df)(__m512d)(A), (int)(imm), (__v2df)_mm_setzero_pd(), \ |
| 1211 | (__mmask8) - 1)) |
| 1212 | |
| 1213 | #define _mm512_mask_extractf64x2_pd(W, U, A, imm) \ |
| 1214 | ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ |
| 1215 | (int)(imm), \ |
| 1216 | (__v2df)(__m128d)(W), \ |
| 1217 | (__mmask8)(U))) |
| 1218 | |
| 1219 | #define _mm512_maskz_extractf64x2_pd(U, A, imm) \ |
| 1220 | ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ |
| 1221 | (int)(imm), \ |
| 1222 | (__v2df)_mm_setzero_pd(), \ |
| 1223 | (__mmask8)(U))) |
| 1224 | |
| 1225 | #define _mm512_extracti32x8_epi32(A, imm) \ |
| 1226 | ((__m256i)__builtin_ia32_extracti32x8_mask( \ |
| 1227 | (__v16si)(__m512i)(A), (int)(imm), (__v8si)_mm256_setzero_si256(), \ |
| 1228 | (__mmask8) - 1)) |
| 1229 | |
| 1230 | #define _mm512_mask_extracti32x8_epi32(W, U, A, imm) \ |
| 1231 | ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ |
| 1232 | (__v8si)(__m256i)(W), \ |
| 1233 | (__mmask8)(U))) |
| 1234 | |
| 1235 | #define _mm512_maskz_extracti32x8_epi32(U, A, imm) \ |
| 1236 | ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ |
| 1237 | (__v8si)_mm256_setzero_si256(), \ |
| 1238 | (__mmask8)(U))) |
| 1239 | |
| 1240 | #define _mm512_extracti64x2_epi64(A, imm) \ |
| 1241 | ((__m128i)__builtin_ia32_extracti64x2_512_mask( \ |
| 1242 | (__v8di)(__m512i)(A), (int)(imm), (__v2di)_mm_setzero_si128(), \ |
| 1243 | (__mmask8) - 1)) |
| 1244 | |
| 1245 | #define _mm512_mask_extracti64x2_epi64(W, U, A, imm) \ |
| 1246 | ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ |
| 1247 | (int)(imm), \ |
| 1248 | (__v2di)(__m128i)(W), \ |
| 1249 | (__mmask8)(U))) |
| 1250 | |
| 1251 | #define _mm512_maskz_extracti64x2_epi64(U, A, imm) \ |
| 1252 | ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ |
| 1253 | (int)(imm), \ |
| 1254 | (__v2di)_mm_setzero_si128(), \ |
| 1255 | (__mmask8)(U))) |
| 1256 | |
| 1257 | #define _mm512_insertf32x8(A, B, imm) \ |
| 1258 | ((__m512)__builtin_ia32_insertf32x8((__v16sf)(__m512)(A), \ |
| 1259 | (__v8sf)(__m256)(B), (int)(imm))) |
| 1260 | |
| 1261 | #define _mm512_mask_insertf32x8(W, U, A, B, imm) \ |
| 1262 | ((__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ |
| 1263 | (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ |
| 1264 | (__v16sf)(__m512)(W))) |
| 1265 | |
| 1266 | #define _mm512_maskz_insertf32x8(U, A, B, imm) \ |
| 1267 | ((__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ |
| 1268 | (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ |
| 1269 | (__v16sf)_mm512_setzero_ps())) |
| 1270 | |
| 1271 | #define _mm512_insertf64x2(A, B, imm) \ |
| 1272 | ((__m512d)__builtin_ia32_insertf64x2_512((__v8df)(__m512d)(A), \ |
| 1273 | (__v2df)(__m128d)(B), (int)(imm))) |
| 1274 | |
| 1275 | #define _mm512_mask_insertf64x2(W, U, A, B, imm) \ |
| 1276 | ((__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ |
| 1277 | (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ |
| 1278 | (__v8df)(__m512d)(W))) |
| 1279 | |
| 1280 | #define _mm512_maskz_insertf64x2(U, A, B, imm) \ |
| 1281 | ((__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ |
| 1282 | (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ |
| 1283 | (__v8df)_mm512_setzero_pd())) |
| 1284 | |
| 1285 | #define _mm512_inserti32x8(A, B, imm) \ |
| 1286 | ((__m512i)__builtin_ia32_inserti32x8((__v16si)(__m512i)(A), \ |
| 1287 | (__v8si)(__m256i)(B), (int)(imm))) |
| 1288 | |
| 1289 | #define _mm512_mask_inserti32x8(W, U, A, B, imm) \ |
| 1290 | ((__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ |
| 1291 | (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ |
| 1292 | (__v16si)(__m512i)(W))) |
| 1293 | |
| 1294 | #define _mm512_maskz_inserti32x8(U, A, B, imm) \ |
| 1295 | ((__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ |
| 1296 | (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ |
| 1297 | (__v16si)_mm512_setzero_si512())) |
| 1298 | |
| 1299 | #define _mm512_inserti64x2(A, B, imm) \ |
| 1300 | ((__m512i)__builtin_ia32_inserti64x2_512((__v8di)(__m512i)(A), \ |
| 1301 | (__v2di)(__m128i)(B), (int)(imm))) |
| 1302 | |
| 1303 | #define _mm512_mask_inserti64x2(W, U, A, B, imm) \ |
| 1304 | ((__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ |
| 1305 | (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ |
| 1306 | (__v8di)(__m512i)(W))) |
| 1307 | |
| 1308 | #define _mm512_maskz_inserti64x2(U, A, B, imm) \ |
| 1309 | ((__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ |
| 1310 | (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ |
| 1311 | (__v8di)_mm512_setzero_si512())) |
| 1312 | |
| 1313 | #define _mm512_mask_fpclass_ps_mask(U, A, imm) \ |
| 1314 | ((__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ |
| 1315 | (int)(imm), (__mmask16)(U))) |
| 1316 | |
| 1317 | #define _mm512_fpclass_ps_mask(A, imm) \ |
| 1318 | ((__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ |
| 1319 | (int)(imm), (__mmask16)-1)) |
| 1320 | |
| 1321 | #define _mm512_mask_fpclass_pd_mask(U, A, imm) \ |
| 1322 | ((__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ |
| 1323 | (__mmask8)(U))) |
| 1324 | |
| 1325 | #define _mm512_fpclass_pd_mask(A, imm) \ |
| 1326 | ((__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ |
| 1327 | (__mmask8)-1)) |
| 1328 | |
| 1329 | #define _mm_fpclass_sd_mask(A, imm) \ |
| 1330 | ((__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ |
| 1331 | (__mmask8)-1)) |
| 1332 | |
| 1333 | #define _mm_mask_fpclass_sd_mask(U, A, imm) \ |
| 1334 | ((__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ |
| 1335 | (__mmask8)(U))) |
| 1336 | |
| 1337 | #define _mm_fpclass_ss_mask(A, imm) \ |
| 1338 | ((__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 1339 | (__mmask8)-1)) |
| 1340 | |
| 1341 | #define _mm_mask_fpclass_ss_mask(U, A, imm) \ |
| 1342 | ((__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 1343 | (__mmask8)(U))) |
| 1344 | |
| 1345 | #undef __DEFAULT_FN_ATTRS512 |
| 1346 | #undef __DEFAULT_FN_ATTRS |
| 1347 | #undef __DEFAULT_FN_ATTRS512_CONSTEXPR |
| 1348 | #undef __DEFAULT_FN_ATTRS_CONSTEXPR |
| 1349 | |
| 1350 | #endif |