| 1 | /*===---- avx512vldqintrin.h - AVX512VL and 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 <avx512vldqintrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif |
| 13 | |
| 14 | #ifndef __AVX512VLDQINTRIN_H |
| 15 | #define __AVX512VLDQINTRIN_H |
| 16 | |
| 17 | /* Define the default attributes for the functions in this file. */ |
| 18 | #define __DEFAULT_FN_ATTRS128 \ |
| 19 | __attribute__((__always_inline__, __nodebug__, \ |
| 20 | __target__("avx512vl,avx512dq"), __min_vector_width__(128))) |
| 21 | #define __DEFAULT_FN_ATTRS256 \ |
| 22 | __attribute__((__always_inline__, __nodebug__, \ |
| 23 | __target__("avx512vl,avx512dq"), __min_vector_width__(256))) |
| 24 | |
| 25 | #if defined(__cplusplus) && (__cplusplus >= 201103L) |
| 26 | #define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 constexpr |
| 27 | #define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 constexpr |
| 28 | #else |
| 29 | #define __DEFAULT_FN_ATTRS256_CONSTEXPR __DEFAULT_FN_ATTRS256 |
| 30 | #define __DEFAULT_FN_ATTRS128_CONSTEXPR __DEFAULT_FN_ATTRS128 |
| 31 | #endif |
| 32 | |
| 33 | static __inline__ __m256i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 34 | _mm256_mullo_epi64(__m256i __A, __m256i __B) { |
| 35 | return (__m256i) ((__v4du) __A * (__v4du) __B); |
| 36 | } |
| 37 | |
| 38 | static __inline__ __m256i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 39 | _mm256_mask_mullo_epi64(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) { |
| 40 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 41 | (__v4di)_mm256_mullo_epi64(__A, __B), |
| 42 | (__v4di)__W); |
| 43 | } |
| 44 | |
| 45 | static __inline__ __m256i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 46 | _mm256_maskz_mullo_epi64(__mmask8 __U, __m256i __A, __m256i __B) { |
| 47 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 48 | (__v4di)_mm256_mullo_epi64(__A, __B), |
| 49 | (__v4di)_mm256_setzero_si256()); |
| 50 | } |
| 51 | |
| 52 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 53 | _mm_mullo_epi64(__m128i __A, __m128i __B) { |
| 54 | return (__m128i) ((__v2du) __A * (__v2du) __B); |
| 55 | } |
| 56 | |
| 57 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 58 | _mm_mask_mullo_epi64(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
| 59 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 60 | (__v2di)_mm_mullo_epi64(__A, __B), |
| 61 | (__v2di)__W); |
| 62 | } |
| 63 | |
| 64 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 65 | _mm_maskz_mullo_epi64(__mmask8 __U, __m128i __A, __m128i __B) { |
| 66 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 67 | (__v2di)_mm_mullo_epi64(__A, __B), |
| 68 | (__v2di)_mm_setzero_si128()); |
| 69 | } |
| 70 | |
| 71 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 72 | _mm256_mask_andnot_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 73 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 74 | (__v4df)_mm256_andnot_pd(__A, __B), |
| 75 | (__v4df)__W); |
| 76 | } |
| 77 | |
| 78 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 79 | _mm256_maskz_andnot_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 80 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 81 | (__v4df)_mm256_andnot_pd(__A, __B), |
| 82 | (__v4df)_mm256_setzero_pd()); |
| 83 | } |
| 84 | |
| 85 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 86 | _mm_mask_andnot_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 87 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 88 | (__v2df)_mm_andnot_pd(__A, __B), |
| 89 | (__v2df)__W); |
| 90 | } |
| 91 | |
| 92 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 93 | _mm_maskz_andnot_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 94 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 95 | (__v2df)_mm_andnot_pd(__A, __B), |
| 96 | (__v2df)_mm_setzero_pd()); |
| 97 | } |
| 98 | |
| 99 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 100 | _mm256_mask_andnot_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 101 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 102 | (__v8sf)_mm256_andnot_ps(__A, __B), |
| 103 | (__v8sf)__W); |
| 104 | } |
| 105 | |
| 106 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 107 | _mm256_maskz_andnot_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 108 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 109 | (__v8sf)_mm256_andnot_ps(__A, __B), |
| 110 | (__v8sf)_mm256_setzero_ps()); |
| 111 | } |
| 112 | |
| 113 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 114 | _mm_mask_andnot_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 115 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 116 | (__v4sf)_mm_andnot_ps(__A, __B), |
| 117 | (__v4sf)__W); |
| 118 | } |
| 119 | |
| 120 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 121 | _mm_maskz_andnot_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 122 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 123 | (__v4sf)_mm_andnot_ps(__A, __B), |
| 124 | (__v4sf)_mm_setzero_ps()); |
| 125 | } |
| 126 | |
| 127 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 128 | _mm256_mask_and_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 129 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 130 | (__v4df)_mm256_and_pd(__A, __B), |
| 131 | (__v4df)__W); |
| 132 | } |
| 133 | |
| 134 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 135 | _mm256_maskz_and_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 136 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 137 | (__v4df)_mm256_and_pd(__A, __B), |
| 138 | (__v4df)_mm256_setzero_pd()); |
| 139 | } |
| 140 | |
| 141 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 142 | _mm_mask_and_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 143 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 144 | (__v2df)_mm_and_pd(__A, __B), |
| 145 | (__v2df)__W); |
| 146 | } |
| 147 | |
| 148 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 149 | _mm_maskz_and_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 150 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 151 | (__v2df)_mm_and_pd(__A, __B), |
| 152 | (__v2df)_mm_setzero_pd()); |
| 153 | } |
| 154 | |
| 155 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 156 | _mm256_mask_and_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 157 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 158 | (__v8sf)_mm256_and_ps(__A, __B), |
| 159 | (__v8sf)__W); |
| 160 | } |
| 161 | |
| 162 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 163 | _mm256_maskz_and_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 164 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 165 | (__v8sf)_mm256_and_ps(__A, __B), |
| 166 | (__v8sf)_mm256_setzero_ps()); |
| 167 | } |
| 168 | |
| 169 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 170 | _mm_mask_and_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 171 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 172 | (__v4sf)_mm_and_ps(__A, __B), |
| 173 | (__v4sf)__W); |
| 174 | } |
| 175 | |
| 176 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 177 | _mm_maskz_and_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 178 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 179 | (__v4sf)_mm_and_ps(__A, __B), |
| 180 | (__v4sf)_mm_setzero_ps()); |
| 181 | } |
| 182 | |
| 183 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 184 | _mm256_mask_xor_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 185 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 186 | (__v4df)_mm256_xor_pd(__A, __B), |
| 187 | (__v4df)__W); |
| 188 | } |
| 189 | |
| 190 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 191 | _mm256_maskz_xor_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 192 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 193 | (__v4df)_mm256_xor_pd(__A, __B), |
| 194 | (__v4df)_mm256_setzero_pd()); |
| 195 | } |
| 196 | |
| 197 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 198 | _mm_mask_xor_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 199 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 200 | (__v2df)_mm_xor_pd(__A, __B), |
| 201 | (__v2df)__W); |
| 202 | } |
| 203 | |
| 204 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 205 | _mm_maskz_xor_pd (__mmask8 __U, __m128d __A, __m128d __B) { |
| 206 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 207 | (__v2df)_mm_xor_pd(__A, __B), |
| 208 | (__v2df)_mm_setzero_pd()); |
| 209 | } |
| 210 | |
| 211 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 212 | _mm256_mask_xor_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 213 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 214 | (__v8sf)_mm256_xor_ps(__A, __B), |
| 215 | (__v8sf)__W); |
| 216 | } |
| 217 | |
| 218 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 219 | _mm256_maskz_xor_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 220 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 221 | (__v8sf)_mm256_xor_ps(__A, __B), |
| 222 | (__v8sf)_mm256_setzero_ps()); |
| 223 | } |
| 224 | |
| 225 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 226 | _mm_mask_xor_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 227 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 228 | (__v4sf)_mm_xor_ps(__A, __B), |
| 229 | (__v4sf)__W); |
| 230 | } |
| 231 | |
| 232 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 233 | _mm_maskz_xor_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 234 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 235 | (__v4sf)_mm_xor_ps(__A, __B), |
| 236 | (__v4sf)_mm_setzero_ps()); |
| 237 | } |
| 238 | |
| 239 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 240 | _mm256_mask_or_pd(__m256d __W, __mmask8 __U, __m256d __A, __m256d __B) { |
| 241 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 242 | (__v4df)_mm256_or_pd(__A, __B), |
| 243 | (__v4df)__W); |
| 244 | } |
| 245 | |
| 246 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
| 247 | _mm256_maskz_or_pd(__mmask8 __U, __m256d __A, __m256d __B) { |
| 248 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 249 | (__v4df)_mm256_or_pd(__A, __B), |
| 250 | (__v4df)_mm256_setzero_pd()); |
| 251 | } |
| 252 | |
| 253 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 254 | _mm_mask_or_pd(__m128d __W, __mmask8 __U, __m128d __A, __m128d __B) { |
| 255 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 256 | (__v2df)_mm_or_pd(__A, __B), |
| 257 | (__v2df)__W); |
| 258 | } |
| 259 | |
| 260 | static __inline__ __m128d __DEFAULT_FN_ATTRS128 |
| 261 | _mm_maskz_or_pd(__mmask8 __U, __m128d __A, __m128d __B) { |
| 262 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 263 | (__v2df)_mm_or_pd(__A, __B), |
| 264 | (__v2df)_mm_setzero_pd()); |
| 265 | } |
| 266 | |
| 267 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 268 | _mm256_mask_or_ps(__m256 __W, __mmask8 __U, __m256 __A, __m256 __B) { |
| 269 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 270 | (__v8sf)_mm256_or_ps(__A, __B), |
| 271 | (__v8sf)__W); |
| 272 | } |
| 273 | |
| 274 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
| 275 | _mm256_maskz_or_ps(__mmask8 __U, __m256 __A, __m256 __B) { |
| 276 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__U, |
| 277 | (__v8sf)_mm256_or_ps(__A, __B), |
| 278 | (__v8sf)_mm256_setzero_ps()); |
| 279 | } |
| 280 | |
| 281 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 282 | _mm_mask_or_ps(__m128 __W, __mmask8 __U, __m128 __A, __m128 __B) { |
| 283 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 284 | (__v4sf)_mm_or_ps(__A, __B), |
| 285 | (__v4sf)__W); |
| 286 | } |
| 287 | |
| 288 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 289 | _mm_maskz_or_ps(__mmask8 __U, __m128 __A, __m128 __B) { |
| 290 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 291 | (__v4sf)_mm_or_ps(__A, __B), |
| 292 | (__v4sf)_mm_setzero_ps()); |
| 293 | } |
| 294 | |
| 295 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 296 | _mm_cvtpd_epi64 (__m128d __A) { |
| 297 | return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A, |
| 298 | (__v2di) _mm_setzero_si128(), |
| 299 | (__mmask8) -1); |
| 300 | } |
| 301 | |
| 302 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 303 | _mm_mask_cvtpd_epi64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 304 | return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A, |
| 305 | (__v2di) __W, |
| 306 | (__mmask8) __U); |
| 307 | } |
| 308 | |
| 309 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 310 | _mm_maskz_cvtpd_epi64 (__mmask8 __U, __m128d __A) { |
| 311 | return (__m128i) __builtin_ia32_cvtpd2qq128_mask ((__v2df) __A, |
| 312 | (__v2di) _mm_setzero_si128(), |
| 313 | (__mmask8) __U); |
| 314 | } |
| 315 | |
| 316 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 317 | _mm256_cvtpd_epi64 (__m256d __A) { |
| 318 | return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A, |
| 319 | (__v4di) _mm256_setzero_si256(), |
| 320 | (__mmask8) -1); |
| 321 | } |
| 322 | |
| 323 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 324 | _mm256_mask_cvtpd_epi64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 325 | return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A, |
| 326 | (__v4di) __W, |
| 327 | (__mmask8) __U); |
| 328 | } |
| 329 | |
| 330 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 331 | _mm256_maskz_cvtpd_epi64 (__mmask8 __U, __m256d __A) { |
| 332 | return (__m256i) __builtin_ia32_cvtpd2qq256_mask ((__v4df) __A, |
| 333 | (__v4di) _mm256_setzero_si256(), |
| 334 | (__mmask8) __U); |
| 335 | } |
| 336 | |
| 337 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 338 | _mm_cvtpd_epu64 (__m128d __A) { |
| 339 | return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A, |
| 340 | (__v2di) _mm_setzero_si128(), |
| 341 | (__mmask8) -1); |
| 342 | } |
| 343 | |
| 344 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 345 | _mm_mask_cvtpd_epu64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 346 | return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A, |
| 347 | (__v2di) __W, |
| 348 | (__mmask8) __U); |
| 349 | } |
| 350 | |
| 351 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 352 | _mm_maskz_cvtpd_epu64 (__mmask8 __U, __m128d __A) { |
| 353 | return (__m128i) __builtin_ia32_cvtpd2uqq128_mask ((__v2df) __A, |
| 354 | (__v2di) _mm_setzero_si128(), |
| 355 | (__mmask8) __U); |
| 356 | } |
| 357 | |
| 358 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 359 | _mm256_cvtpd_epu64 (__m256d __A) { |
| 360 | return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A, |
| 361 | (__v4di) _mm256_setzero_si256(), |
| 362 | (__mmask8) -1); |
| 363 | } |
| 364 | |
| 365 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 366 | _mm256_mask_cvtpd_epu64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 367 | return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A, |
| 368 | (__v4di) __W, |
| 369 | (__mmask8) __U); |
| 370 | } |
| 371 | |
| 372 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 373 | _mm256_maskz_cvtpd_epu64 (__mmask8 __U, __m256d __A) { |
| 374 | return (__m256i) __builtin_ia32_cvtpd2uqq256_mask ((__v4df) __A, |
| 375 | (__v4di) _mm256_setzero_si256(), |
| 376 | (__mmask8) __U); |
| 377 | } |
| 378 | |
| 379 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 380 | _mm_cvtps_epi64 (__m128 __A) { |
| 381 | return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A, |
| 382 | (__v2di) _mm_setzero_si128(), |
| 383 | (__mmask8) -1); |
| 384 | } |
| 385 | |
| 386 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 387 | _mm_mask_cvtps_epi64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 388 | return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A, |
| 389 | (__v2di) __W, |
| 390 | (__mmask8) __U); |
| 391 | } |
| 392 | |
| 393 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 394 | _mm_maskz_cvtps_epi64 (__mmask8 __U, __m128 __A) { |
| 395 | return (__m128i) __builtin_ia32_cvtps2qq128_mask ((__v4sf) __A, |
| 396 | (__v2di) _mm_setzero_si128(), |
| 397 | (__mmask8) __U); |
| 398 | } |
| 399 | |
| 400 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 401 | _mm256_cvtps_epi64 (__m128 __A) { |
| 402 | return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A, |
| 403 | (__v4di) _mm256_setzero_si256(), |
| 404 | (__mmask8) -1); |
| 405 | } |
| 406 | |
| 407 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 408 | _mm256_mask_cvtps_epi64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 409 | return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A, |
| 410 | (__v4di) __W, |
| 411 | (__mmask8) __U); |
| 412 | } |
| 413 | |
| 414 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 415 | _mm256_maskz_cvtps_epi64 (__mmask8 __U, __m128 __A) { |
| 416 | return (__m256i) __builtin_ia32_cvtps2qq256_mask ((__v4sf) __A, |
| 417 | (__v4di) _mm256_setzero_si256(), |
| 418 | (__mmask8) __U); |
| 419 | } |
| 420 | |
| 421 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 422 | _mm_cvtps_epu64 (__m128 __A) { |
| 423 | return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A, |
| 424 | (__v2di) _mm_setzero_si128(), |
| 425 | (__mmask8) -1); |
| 426 | } |
| 427 | |
| 428 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 429 | _mm_mask_cvtps_epu64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 430 | return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A, |
| 431 | (__v2di) __W, |
| 432 | (__mmask8) __U); |
| 433 | } |
| 434 | |
| 435 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 436 | _mm_maskz_cvtps_epu64 (__mmask8 __U, __m128 __A) { |
| 437 | return (__m128i) __builtin_ia32_cvtps2uqq128_mask ((__v4sf) __A, |
| 438 | (__v2di) _mm_setzero_si128(), |
| 439 | (__mmask8) __U); |
| 440 | } |
| 441 | |
| 442 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 443 | _mm256_cvtps_epu64 (__m128 __A) { |
| 444 | return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A, |
| 445 | (__v4di) _mm256_setzero_si256(), |
| 446 | (__mmask8) -1); |
| 447 | } |
| 448 | |
| 449 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 450 | _mm256_mask_cvtps_epu64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 451 | return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A, |
| 452 | (__v4di) __W, |
| 453 | (__mmask8) __U); |
| 454 | } |
| 455 | |
| 456 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 457 | _mm256_maskz_cvtps_epu64 (__mmask8 __U, __m128 __A) { |
| 458 | return (__m256i) __builtin_ia32_cvtps2uqq256_mask ((__v4sf) __A, |
| 459 | (__v4di) _mm256_setzero_si256(), |
| 460 | (__mmask8) __U); |
| 461 | } |
| 462 | |
| 463 | static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 464 | _mm_cvtepi64_pd(__m128i __A) { |
| 465 | return (__m128d)__builtin_convertvector((__v2di)__A, __v2df); |
| 466 | } |
| 467 | |
| 468 | static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 469 | _mm_mask_cvtepi64_pd(__m128d __W, __mmask8 __U, __m128i __A) { |
| 470 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 471 | (__v2df)_mm_cvtepi64_pd(__A), |
| 472 | (__v2df)__W); |
| 473 | } |
| 474 | |
| 475 | static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 476 | _mm_maskz_cvtepi64_pd(__mmask8 __U, __m128i __A) { |
| 477 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 478 | (__v2df)_mm_cvtepi64_pd(__A), |
| 479 | (__v2df)_mm_setzero_pd()); |
| 480 | } |
| 481 | |
| 482 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 483 | _mm256_cvtepi64_pd(__m256i __A) { |
| 484 | return (__m256d)__builtin_convertvector((__v4di)__A, __v4df); |
| 485 | } |
| 486 | |
| 487 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 488 | _mm256_mask_cvtepi64_pd(__m256d __W, __mmask8 __U, __m256i __A) { |
| 489 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 490 | (__v4df)_mm256_cvtepi64_pd(__A), |
| 491 | (__v4df)__W); |
| 492 | } |
| 493 | |
| 494 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 495 | _mm256_maskz_cvtepi64_pd(__mmask8 __U, __m256i __A) { |
| 496 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 497 | (__v4df)_mm256_cvtepi64_pd(__A), |
| 498 | (__v4df)_mm256_setzero_pd()); |
| 499 | } |
| 500 | |
| 501 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 502 | _mm_cvtepi64_ps (__m128i __A) { |
| 503 | return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A, |
| 504 | (__v4sf) _mm_setzero_ps(), |
| 505 | (__mmask8) -1); |
| 506 | } |
| 507 | |
| 508 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 509 | _mm_mask_cvtepi64_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 510 | return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A, |
| 511 | (__v4sf) __W, |
| 512 | (__mmask8) __U); |
| 513 | } |
| 514 | |
| 515 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 516 | _mm_maskz_cvtepi64_ps (__mmask8 __U, __m128i __A) { |
| 517 | return (__m128) __builtin_ia32_cvtqq2ps128_mask ((__v2di) __A, |
| 518 | (__v4sf) _mm_setzero_ps(), |
| 519 | (__mmask8) __U); |
| 520 | } |
| 521 | |
| 522 | static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 523 | _mm256_cvtepi64_ps(__m256i __A) { |
| 524 | return (__m128)__builtin_convertvector((__v4di)__A, __v4sf); |
| 525 | } |
| 526 | |
| 527 | static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 528 | _mm256_mask_cvtepi64_ps(__m128 __W, __mmask8 __U, __m256i __A) { |
| 529 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 530 | (__v4sf)_mm256_cvtepi64_ps(__A), |
| 531 | (__v4sf)__W); |
| 532 | } |
| 533 | |
| 534 | static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 535 | _mm256_maskz_cvtepi64_ps(__mmask8 __U, __m256i __A) { |
| 536 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 537 | (__v4sf)_mm256_cvtepi64_ps(__A), |
| 538 | (__v4sf)_mm_setzero_ps()); |
| 539 | } |
| 540 | |
| 541 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 542 | _mm_cvttpd_epi64 (__m128d __A) { |
| 543 | return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A, |
| 544 | (__v2di) _mm_setzero_si128(), |
| 545 | (__mmask8) -1); |
| 546 | } |
| 547 | |
| 548 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 549 | _mm_mask_cvttpd_epi64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 550 | return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A, |
| 551 | (__v2di) __W, |
| 552 | (__mmask8) __U); |
| 553 | } |
| 554 | |
| 555 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 556 | _mm_maskz_cvttpd_epi64 (__mmask8 __U, __m128d __A) { |
| 557 | return (__m128i) __builtin_ia32_cvttpd2qq128_mask ((__v2df) __A, |
| 558 | (__v2di) _mm_setzero_si128(), |
| 559 | (__mmask8) __U); |
| 560 | } |
| 561 | |
| 562 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 563 | _mm256_cvttpd_epi64 (__m256d __A) { |
| 564 | return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A, |
| 565 | (__v4di) _mm256_setzero_si256(), |
| 566 | (__mmask8) -1); |
| 567 | } |
| 568 | |
| 569 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 570 | _mm256_mask_cvttpd_epi64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 571 | return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A, |
| 572 | (__v4di) __W, |
| 573 | (__mmask8) __U); |
| 574 | } |
| 575 | |
| 576 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 577 | _mm256_maskz_cvttpd_epi64 (__mmask8 __U, __m256d __A) { |
| 578 | return (__m256i) __builtin_ia32_cvttpd2qq256_mask ((__v4df) __A, |
| 579 | (__v4di) _mm256_setzero_si256(), |
| 580 | (__mmask8) __U); |
| 581 | } |
| 582 | |
| 583 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 584 | _mm_cvttpd_epu64 (__m128d __A) { |
| 585 | return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A, |
| 586 | (__v2di) _mm_setzero_si128(), |
| 587 | (__mmask8) -1); |
| 588 | } |
| 589 | |
| 590 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 591 | _mm_mask_cvttpd_epu64 (__m128i __W, __mmask8 __U, __m128d __A) { |
| 592 | return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A, |
| 593 | (__v2di) __W, |
| 594 | (__mmask8) __U); |
| 595 | } |
| 596 | |
| 597 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 598 | _mm_maskz_cvttpd_epu64 (__mmask8 __U, __m128d __A) { |
| 599 | return (__m128i) __builtin_ia32_cvttpd2uqq128_mask ((__v2df) __A, |
| 600 | (__v2di) _mm_setzero_si128(), |
| 601 | (__mmask8) __U); |
| 602 | } |
| 603 | |
| 604 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 605 | _mm256_cvttpd_epu64 (__m256d __A) { |
| 606 | return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A, |
| 607 | (__v4di) _mm256_setzero_si256(), |
| 608 | (__mmask8) -1); |
| 609 | } |
| 610 | |
| 611 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 612 | _mm256_mask_cvttpd_epu64 (__m256i __W, __mmask8 __U, __m256d __A) { |
| 613 | return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A, |
| 614 | (__v4di) __W, |
| 615 | (__mmask8) __U); |
| 616 | } |
| 617 | |
| 618 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 619 | _mm256_maskz_cvttpd_epu64 (__mmask8 __U, __m256d __A) { |
| 620 | return (__m256i) __builtin_ia32_cvttpd2uqq256_mask ((__v4df) __A, |
| 621 | (__v4di) _mm256_setzero_si256(), |
| 622 | (__mmask8) __U); |
| 623 | } |
| 624 | |
| 625 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 626 | _mm_cvttps_epi64 (__m128 __A) { |
| 627 | return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A, |
| 628 | (__v2di) _mm_setzero_si128(), |
| 629 | (__mmask8) -1); |
| 630 | } |
| 631 | |
| 632 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 633 | _mm_mask_cvttps_epi64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 634 | return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A, |
| 635 | (__v2di) __W, |
| 636 | (__mmask8) __U); |
| 637 | } |
| 638 | |
| 639 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 640 | _mm_maskz_cvttps_epi64 (__mmask8 __U, __m128 __A) { |
| 641 | return (__m128i) __builtin_ia32_cvttps2qq128_mask ((__v4sf) __A, |
| 642 | (__v2di) _mm_setzero_si128(), |
| 643 | (__mmask8) __U); |
| 644 | } |
| 645 | |
| 646 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 647 | _mm256_cvttps_epi64 (__m128 __A) { |
| 648 | return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A, |
| 649 | (__v4di) _mm256_setzero_si256(), |
| 650 | (__mmask8) -1); |
| 651 | } |
| 652 | |
| 653 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 654 | _mm256_mask_cvttps_epi64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 655 | return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A, |
| 656 | (__v4di) __W, |
| 657 | (__mmask8) __U); |
| 658 | } |
| 659 | |
| 660 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 661 | _mm256_maskz_cvttps_epi64 (__mmask8 __U, __m128 __A) { |
| 662 | return (__m256i) __builtin_ia32_cvttps2qq256_mask ((__v4sf) __A, |
| 663 | (__v4di) _mm256_setzero_si256(), |
| 664 | (__mmask8) __U); |
| 665 | } |
| 666 | |
| 667 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 668 | _mm_cvttps_epu64 (__m128 __A) { |
| 669 | return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A, |
| 670 | (__v2di) _mm_setzero_si128(), |
| 671 | (__mmask8) -1); |
| 672 | } |
| 673 | |
| 674 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 675 | _mm_mask_cvttps_epu64 (__m128i __W, __mmask8 __U, __m128 __A) { |
| 676 | return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A, |
| 677 | (__v2di) __W, |
| 678 | (__mmask8) __U); |
| 679 | } |
| 680 | |
| 681 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 682 | _mm_maskz_cvttps_epu64 (__mmask8 __U, __m128 __A) { |
| 683 | return (__m128i) __builtin_ia32_cvttps2uqq128_mask ((__v4sf) __A, |
| 684 | (__v2di) _mm_setzero_si128(), |
| 685 | (__mmask8) __U); |
| 686 | } |
| 687 | |
| 688 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 689 | _mm256_cvttps_epu64 (__m128 __A) { |
| 690 | return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A, |
| 691 | (__v4di) _mm256_setzero_si256(), |
| 692 | (__mmask8) -1); |
| 693 | } |
| 694 | |
| 695 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 696 | _mm256_mask_cvttps_epu64 (__m256i __W, __mmask8 __U, __m128 __A) { |
| 697 | return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A, |
| 698 | (__v4di) __W, |
| 699 | (__mmask8) __U); |
| 700 | } |
| 701 | |
| 702 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 703 | _mm256_maskz_cvttps_epu64 (__mmask8 __U, __m128 __A) { |
| 704 | return (__m256i) __builtin_ia32_cvttps2uqq256_mask ((__v4sf) __A, |
| 705 | (__v4di) _mm256_setzero_si256(), |
| 706 | (__mmask8) __U); |
| 707 | } |
| 708 | |
| 709 | static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 710 | _mm_cvtepu64_pd(__m128i __A) { |
| 711 | return (__m128d)__builtin_convertvector((__v2du)__A, __v2df); |
| 712 | } |
| 713 | |
| 714 | static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 715 | _mm_mask_cvtepu64_pd(__m128d __W, __mmask8 __U, __m128i __A) { |
| 716 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 717 | (__v2df)_mm_cvtepu64_pd(__A), |
| 718 | (__v2df)__W); |
| 719 | } |
| 720 | |
| 721 | static __inline__ __m128d __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 722 | _mm_maskz_cvtepu64_pd(__mmask8 __U, __m128i __A) { |
| 723 | return (__m128d)__builtin_ia32_selectpd_128((__mmask8)__U, |
| 724 | (__v2df)_mm_cvtepu64_pd(__A), |
| 725 | (__v2df)_mm_setzero_pd()); |
| 726 | } |
| 727 | |
| 728 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 729 | _mm256_cvtepu64_pd(__m256i __A) { |
| 730 | return (__m256d)__builtin_convertvector((__v4du)__A, __v4df); |
| 731 | } |
| 732 | |
| 733 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 734 | _mm256_mask_cvtepu64_pd(__m256d __W, __mmask8 __U, __m256i __A) { |
| 735 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 736 | (__v4df)_mm256_cvtepu64_pd(__A), |
| 737 | (__v4df)__W); |
| 738 | } |
| 739 | |
| 740 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 741 | _mm256_maskz_cvtepu64_pd(__mmask8 __U, __m256i __A) { |
| 742 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__U, |
| 743 | (__v4df)_mm256_cvtepu64_pd(__A), |
| 744 | (__v4df)_mm256_setzero_pd()); |
| 745 | } |
| 746 | |
| 747 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 748 | _mm_cvtepu64_ps (__m128i __A) { |
| 749 | return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A, |
| 750 | (__v4sf) _mm_setzero_ps(), |
| 751 | (__mmask8) -1); |
| 752 | } |
| 753 | |
| 754 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 755 | _mm_mask_cvtepu64_ps (__m128 __W, __mmask8 __U, __m128i __A) { |
| 756 | return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A, |
| 757 | (__v4sf) __W, |
| 758 | (__mmask8) __U); |
| 759 | } |
| 760 | |
| 761 | static __inline__ __m128 __DEFAULT_FN_ATTRS128 |
| 762 | _mm_maskz_cvtepu64_ps (__mmask8 __U, __m128i __A) { |
| 763 | return (__m128) __builtin_ia32_cvtuqq2ps128_mask ((__v2di) __A, |
| 764 | (__v4sf) _mm_setzero_ps(), |
| 765 | (__mmask8) __U); |
| 766 | } |
| 767 | |
| 768 | static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 769 | _mm256_cvtepu64_ps(__m256i __A) { |
| 770 | return (__m128)__builtin_convertvector((__v4du)__A, __v4sf); |
| 771 | } |
| 772 | |
| 773 | static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 774 | _mm256_mask_cvtepu64_ps(__m128 __W, __mmask8 __U, __m256i __A) { |
| 775 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 776 | (__v4sf)_mm256_cvtepu64_ps(__A), |
| 777 | (__v4sf)__W); |
| 778 | } |
| 779 | |
| 780 | static __inline__ __m128 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 781 | _mm256_maskz_cvtepu64_ps(__mmask8 __U, __m256i __A) { |
| 782 | return (__m128)__builtin_ia32_selectps_128((__mmask8)__U, |
| 783 | (__v4sf)_mm256_cvtepu64_ps(__A), |
| 784 | (__v4sf)_mm_setzero_ps()); |
| 785 | } |
| 786 | |
| 787 | #define _mm_range_pd(A, B, C) \ |
| 788 | ((__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \ |
| 789 | (__v2df)(__m128d)(B), (int)(C), \ |
| 790 | (__v2df)_mm_setzero_pd(), \ |
| 791 | (__mmask8)-1)) |
| 792 | |
| 793 | #define _mm_mask_range_pd(W, U, A, B, C) \ |
| 794 | ((__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \ |
| 795 | (__v2df)(__m128d)(B), (int)(C), \ |
| 796 | (__v2df)(__m128d)(W), \ |
| 797 | (__mmask8)(U))) |
| 798 | |
| 799 | #define _mm_maskz_range_pd(U, A, B, C) \ |
| 800 | ((__m128d)__builtin_ia32_rangepd128_mask((__v2df)(__m128d)(A), \ |
| 801 | (__v2df)(__m128d)(B), (int)(C), \ |
| 802 | (__v2df)_mm_setzero_pd(), \ |
| 803 | (__mmask8)(U))) |
| 804 | |
| 805 | #define _mm256_range_pd(A, B, C) \ |
| 806 | ((__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \ |
| 807 | (__v4df)(__m256d)(B), (int)(C), \ |
| 808 | (__v4df)_mm256_setzero_pd(), \ |
| 809 | (__mmask8)-1)) |
| 810 | |
| 811 | #define _mm256_mask_range_pd(W, U, A, B, C) \ |
| 812 | ((__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \ |
| 813 | (__v4df)(__m256d)(B), (int)(C), \ |
| 814 | (__v4df)(__m256d)(W), \ |
| 815 | (__mmask8)(U))) |
| 816 | |
| 817 | #define _mm256_maskz_range_pd(U, A, B, C) \ |
| 818 | ((__m256d)__builtin_ia32_rangepd256_mask((__v4df)(__m256d)(A), \ |
| 819 | (__v4df)(__m256d)(B), (int)(C), \ |
| 820 | (__v4df)_mm256_setzero_pd(), \ |
| 821 | (__mmask8)(U))) |
| 822 | |
| 823 | #define _mm_range_ps(A, B, C) \ |
| 824 | ((__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \ |
| 825 | (__v4sf)(__m128)(B), (int)(C), \ |
| 826 | (__v4sf)_mm_setzero_ps(), \ |
| 827 | (__mmask8)-1)) |
| 828 | |
| 829 | #define _mm_mask_range_ps(W, U, A, B, C) \ |
| 830 | ((__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \ |
| 831 | (__v4sf)(__m128)(B), (int)(C), \ |
| 832 | (__v4sf)(__m128)(W), (__mmask8)(U))) |
| 833 | |
| 834 | #define _mm_maskz_range_ps(U, A, B, C) \ |
| 835 | ((__m128)__builtin_ia32_rangeps128_mask((__v4sf)(__m128)(A), \ |
| 836 | (__v4sf)(__m128)(B), (int)(C), \ |
| 837 | (__v4sf)_mm_setzero_ps(), \ |
| 838 | (__mmask8)(U))) |
| 839 | |
| 840 | #define _mm256_range_ps(A, B, C) \ |
| 841 | ((__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \ |
| 842 | (__v8sf)(__m256)(B), (int)(C), \ |
| 843 | (__v8sf)_mm256_setzero_ps(), \ |
| 844 | (__mmask8)-1)) |
| 845 | |
| 846 | #define _mm256_mask_range_ps(W, U, A, B, C) \ |
| 847 | ((__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \ |
| 848 | (__v8sf)(__m256)(B), (int)(C), \ |
| 849 | (__v8sf)(__m256)(W), (__mmask8)(U))) |
| 850 | |
| 851 | #define _mm256_maskz_range_ps(U, A, B, C) \ |
| 852 | ((__m256)__builtin_ia32_rangeps256_mask((__v8sf)(__m256)(A), \ |
| 853 | (__v8sf)(__m256)(B), (int)(C), \ |
| 854 | (__v8sf)_mm256_setzero_ps(), \ |
| 855 | (__mmask8)(U))) |
| 856 | |
| 857 | #define _mm_reduce_pd(A, B) \ |
| 858 | ((__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \ |
| 859 | (__v2df)_mm_setzero_pd(), \ |
| 860 | (__mmask8)-1)) |
| 861 | |
| 862 | #define _mm_mask_reduce_pd(W, U, A, B) \ |
| 863 | ((__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \ |
| 864 | (__v2df)(__m128d)(W), \ |
| 865 | (__mmask8)(U))) |
| 866 | |
| 867 | #define _mm_maskz_reduce_pd(U, A, B) \ |
| 868 | ((__m128d)__builtin_ia32_reducepd128_mask((__v2df)(__m128d)(A), (int)(B), \ |
| 869 | (__v2df)_mm_setzero_pd(), \ |
| 870 | (__mmask8)(U))) |
| 871 | |
| 872 | #define _mm256_reduce_pd(A, B) \ |
| 873 | ((__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \ |
| 874 | (__v4df)_mm256_setzero_pd(), \ |
| 875 | (__mmask8)-1)) |
| 876 | |
| 877 | #define _mm256_mask_reduce_pd(W, U, A, B) \ |
| 878 | ((__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \ |
| 879 | (__v4df)(__m256d)(W), \ |
| 880 | (__mmask8)(U))) |
| 881 | |
| 882 | #define _mm256_maskz_reduce_pd(U, A, B) \ |
| 883 | ((__m256d)__builtin_ia32_reducepd256_mask((__v4df)(__m256d)(A), (int)(B), \ |
| 884 | (__v4df)_mm256_setzero_pd(), \ |
| 885 | (__mmask8)(U))) |
| 886 | |
| 887 | #define _mm_reduce_ps(A, B) \ |
| 888 | ((__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \ |
| 889 | (__v4sf)_mm_setzero_ps(), \ |
| 890 | (__mmask8)-1)) |
| 891 | |
| 892 | #define _mm_mask_reduce_ps(W, U, A, B) \ |
| 893 | ((__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \ |
| 894 | (__v4sf)(__m128)(W), \ |
| 895 | (__mmask8)(U))) |
| 896 | |
| 897 | #define _mm_maskz_reduce_ps(U, A, B) \ |
| 898 | ((__m128)__builtin_ia32_reduceps128_mask((__v4sf)(__m128)(A), (int)(B), \ |
| 899 | (__v4sf)_mm_setzero_ps(), \ |
| 900 | (__mmask8)(U))) |
| 901 | |
| 902 | #define _mm256_reduce_ps(A, B) \ |
| 903 | ((__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \ |
| 904 | (__v8sf)_mm256_setzero_ps(), \ |
| 905 | (__mmask8)-1)) |
| 906 | |
| 907 | #define _mm256_mask_reduce_ps(W, U, A, B) \ |
| 908 | ((__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \ |
| 909 | (__v8sf)(__m256)(W), \ |
| 910 | (__mmask8)(U))) |
| 911 | |
| 912 | #define _mm256_maskz_reduce_ps(U, A, B) \ |
| 913 | ((__m256)__builtin_ia32_reduceps256_mask((__v8sf)(__m256)(A), (int)(B), \ |
| 914 | (__v8sf)_mm256_setzero_ps(), \ |
| 915 | (__mmask8)(U))) |
| 916 | |
| 917 | static __inline__ __mmask8 |
| 918 | __DEFAULT_FN_ATTRS128_CONSTEXPR _mm_movepi32_mask(__m128i __A) { |
| 919 | return (__mmask8) __builtin_ia32_cvtd2mask128 ((__v4si) __A); |
| 920 | } |
| 921 | |
| 922 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 923 | _mm256_movepi32_mask(__m256i __A) { |
| 924 | return (__mmask8) __builtin_ia32_cvtd2mask256 ((__v8si) __A); |
| 925 | } |
| 926 | |
| 927 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 928 | _mm_movm_epi32(__mmask8 __A) { |
| 929 | return (__m128i) __builtin_ia32_cvtmask2d128 (__A); |
| 930 | } |
| 931 | |
| 932 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 933 | _mm256_movm_epi32(__mmask8 __A) { |
| 934 | return (__m256i) __builtin_ia32_cvtmask2d256 (__A); |
| 935 | } |
| 936 | |
| 937 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 938 | _mm_movm_epi64(__mmask8 __A) { |
| 939 | return (__m128i) __builtin_ia32_cvtmask2q128 (__A); |
| 940 | } |
| 941 | |
| 942 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 943 | _mm256_movm_epi64(__mmask8 __A) { |
| 944 | return (__m256i) __builtin_ia32_cvtmask2q256 (__A); |
| 945 | } |
| 946 | |
| 947 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 948 | _mm_movepi64_mask(__m128i __A) { |
| 949 | return (__mmask8) __builtin_ia32_cvtq2mask128 ((__v2di) __A); |
| 950 | } |
| 951 | |
| 952 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 953 | _mm256_movepi64_mask(__m256i __A) { |
| 954 | return (__mmask8) __builtin_ia32_cvtq2mask256 ((__v4di) __A); |
| 955 | } |
| 956 | |
| 957 | static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 958 | _mm256_broadcast_f32x2(__m128 __A) { |
| 959 | return (__m256)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A, |
| 960 | 0, 1, 0, 1, 0, 1, 0, 1); |
| 961 | } |
| 962 | |
| 963 | static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 964 | _mm256_mask_broadcast_f32x2(__m256 __O, __mmask8 __M, __m128 __A) { |
| 965 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__M, |
| 966 | (__v8sf)_mm256_broadcast_f32x2(__A), |
| 967 | (__v8sf)__O); |
| 968 | } |
| 969 | |
| 970 | static __inline__ __m256 __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 971 | _mm256_maskz_broadcast_f32x2(__mmask8 __M, __m128 __A) { |
| 972 | return (__m256)__builtin_ia32_selectps_256((__mmask8)__M, |
| 973 | (__v8sf)_mm256_broadcast_f32x2(__A), |
| 974 | (__v8sf)_mm256_setzero_ps()); |
| 975 | } |
| 976 | |
| 977 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 978 | _mm256_broadcast_f64x2(__m128d __A) { |
| 979 | return (__m256d)__builtin_shufflevector((__v2df)__A, (__v2df)__A, |
| 980 | 0, 1, 0, 1); |
| 981 | } |
| 982 | |
| 983 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 984 | _mm256_mask_broadcast_f64x2(__m256d __O, __mmask8 __M, __m128d __A) { |
| 985 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__M, |
| 986 | (__v4df)_mm256_broadcast_f64x2(__A), |
| 987 | (__v4df)__O); |
| 988 | } |
| 989 | |
| 990 | static __inline__ __m256d __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 991 | _mm256_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A) { |
| 992 | return (__m256d)__builtin_ia32_selectpd_256((__mmask8)__M, |
| 993 | (__v4df)_mm256_broadcast_f64x2(__A), |
| 994 | (__v4df)_mm256_setzero_pd()); |
| 995 | } |
| 996 | |
| 997 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 998 | _mm_broadcast_i32x2(__m128i __A) { |
| 999 | return (__m128i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
| 1000 | 0, 1, 0, 1); |
| 1001 | } |
| 1002 | |
| 1003 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 1004 | _mm_mask_broadcast_i32x2(__m128i __O, __mmask8 __M, __m128i __A) { |
| 1005 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 1006 | (__v4si)_mm_broadcast_i32x2(__A), |
| 1007 | (__v4si)__O); |
| 1008 | } |
| 1009 | |
| 1010 | static __inline__ __m128i __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 1011 | _mm_maskz_broadcast_i32x2(__mmask8 __M, __m128i __A) { |
| 1012 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__M, |
| 1013 | (__v4si)_mm_broadcast_i32x2(__A), |
| 1014 | (__v4si)_mm_setzero_si128()); |
| 1015 | } |
| 1016 | |
| 1017 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1018 | _mm256_broadcast_i32x2(__m128i __A) { |
| 1019 | return (__m256i)__builtin_shufflevector((__v4si)__A, (__v4si)__A, |
| 1020 | 0, 1, 0, 1, 0, 1, 0, 1); |
| 1021 | } |
| 1022 | |
| 1023 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1024 | _mm256_mask_broadcast_i32x2(__m256i __O, __mmask8 __M, __m128i __A) { |
| 1025 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 1026 | (__v8si)_mm256_broadcast_i32x2(__A), |
| 1027 | (__v8si)__O); |
| 1028 | } |
| 1029 | |
| 1030 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1031 | _mm256_maskz_broadcast_i32x2(__mmask8 __M, __m128i __A) { |
| 1032 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__M, |
| 1033 | (__v8si)_mm256_broadcast_i32x2(__A), |
| 1034 | (__v8si)_mm256_setzero_si256()); |
| 1035 | } |
| 1036 | |
| 1037 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1038 | _mm256_broadcast_i64x2(__m128i __A) { |
| 1039 | return (__m256i)__builtin_shufflevector((__v2di)__A, (__v2di)__A, |
| 1040 | 0, 1, 0, 1); |
| 1041 | } |
| 1042 | |
| 1043 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1044 | _mm256_mask_broadcast_i64x2(__m256i __O, __mmask8 __M, __m128i __A) { |
| 1045 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 1046 | (__v4di)_mm256_broadcast_i64x2(__A), |
| 1047 | (__v4di)__O); |
| 1048 | } |
| 1049 | |
| 1050 | static __inline__ __m256i __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1051 | _mm256_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A) { |
| 1052 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__M, |
| 1053 | (__v4di)_mm256_broadcast_i64x2(__A), |
| 1054 | (__v4di)_mm256_setzero_si256()); |
| 1055 | } |
| 1056 | |
| 1057 | #define _mm256_extractf64x2_pd(A, imm) \ |
| 1058 | ((__m128d)__builtin_ia32_extractf64x2_256_mask( \ |
| 1059 | (__v4df)(__m256d)(A), (int)(imm), (__v2df)_mm_setzero_pd(), \ |
| 1060 | (__mmask8) - 1)) |
| 1061 | |
| 1062 | #define _mm256_mask_extractf64x2_pd(W, U, A, imm) \ |
| 1063 | ((__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \ |
| 1064 | (int)(imm), \ |
| 1065 | (__v2df)(__m128d)(W), \ |
| 1066 | (__mmask8)(U))) |
| 1067 | |
| 1068 | #define _mm256_maskz_extractf64x2_pd(U, A, imm) \ |
| 1069 | ((__m128d)__builtin_ia32_extractf64x2_256_mask((__v4df)(__m256d)(A), \ |
| 1070 | (int)(imm), \ |
| 1071 | (__v2df)_mm_setzero_pd(), \ |
| 1072 | (__mmask8)(U))) |
| 1073 | |
| 1074 | #define _mm256_extracti64x2_epi64(A, imm) \ |
| 1075 | ((__m128i)__builtin_ia32_extracti64x2_256_mask( \ |
| 1076 | (__v4di)(__m256i)(A), (int)(imm), (__v2di)_mm_setzero_si128(), \ |
| 1077 | (__mmask8) - 1)) |
| 1078 | |
| 1079 | #define _mm256_mask_extracti64x2_epi64(W, U, A, imm) \ |
| 1080 | ((__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \ |
| 1081 | (int)(imm), \ |
| 1082 | (__v2di)(__m128i)(W), \ |
| 1083 | (__mmask8)(U))) |
| 1084 | |
| 1085 | #define _mm256_maskz_extracti64x2_epi64(U, A, imm) \ |
| 1086 | ((__m128i)__builtin_ia32_extracti64x2_256_mask((__v4di)(__m256i)(A), \ |
| 1087 | (int)(imm), \ |
| 1088 | (__v2di)_mm_setzero_si128(), \ |
| 1089 | (__mmask8)(U))) |
| 1090 | |
| 1091 | #define _mm256_insertf64x2(A, B, imm) \ |
| 1092 | ((__m256d)__builtin_ia32_insertf64x2_256((__v4df)(__m256d)(A), \ |
| 1093 | (__v2df)(__m128d)(B), (int)(imm))) |
| 1094 | |
| 1095 | #define _mm256_mask_insertf64x2(W, U, A, B, imm) \ |
| 1096 | ((__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 1097 | (__v4df)_mm256_insertf64x2((A), (B), (imm)), \ |
| 1098 | (__v4df)(__m256d)(W))) |
| 1099 | |
| 1100 | #define _mm256_maskz_insertf64x2(U, A, B, imm) \ |
| 1101 | ((__m256d)__builtin_ia32_selectpd_256((__mmask8)(U), \ |
| 1102 | (__v4df)_mm256_insertf64x2((A), (B), (imm)), \ |
| 1103 | (__v4df)_mm256_setzero_pd())) |
| 1104 | |
| 1105 | #define _mm256_inserti64x2(A, B, imm) \ |
| 1106 | ((__m256i)__builtin_ia32_inserti64x2_256((__v4di)(__m256i)(A), \ |
| 1107 | (__v2di)(__m128i)(B), (int)(imm))) |
| 1108 | |
| 1109 | #define _mm256_mask_inserti64x2(W, U, A, B, imm) \ |
| 1110 | ((__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 1111 | (__v4di)_mm256_inserti64x2((A), (B), (imm)), \ |
| 1112 | (__v4di)(__m256i)(W))) |
| 1113 | |
| 1114 | #define _mm256_maskz_inserti64x2(U, A, B, imm) \ |
| 1115 | ((__m256i)__builtin_ia32_selectq_256((__mmask8)(U), \ |
| 1116 | (__v4di)_mm256_inserti64x2((A), (B), (imm)), \ |
| 1117 | (__v4di)_mm256_setzero_si256())) |
| 1118 | |
| 1119 | #define _mm_mask_fpclass_pd_mask(U, A, imm) \ |
| 1120 | ((__mmask8)__builtin_ia32_fpclasspd128_mask((__v2df)(__m128d)(A), (int)(imm), \ |
| 1121 | (__mmask8)(U))) |
| 1122 | |
| 1123 | #define _mm_fpclass_pd_mask(A, imm) \ |
| 1124 | ((__mmask8)__builtin_ia32_fpclasspd128_mask((__v2df)(__m128d)(A), (int)(imm), \ |
| 1125 | (__mmask8)-1)) |
| 1126 | |
| 1127 | #define _mm256_mask_fpclass_pd_mask(U, A, imm) \ |
| 1128 | ((__mmask8)__builtin_ia32_fpclasspd256_mask((__v4df)(__m256d)(A), (int)(imm), \ |
| 1129 | (__mmask8)(U))) |
| 1130 | |
| 1131 | #define _mm256_fpclass_pd_mask(A, imm) \ |
| 1132 | ((__mmask8)__builtin_ia32_fpclasspd256_mask((__v4df)(__m256d)(A), (int)(imm), \ |
| 1133 | (__mmask8)-1)) |
| 1134 | |
| 1135 | #define _mm_mask_fpclass_ps_mask(U, A, imm) \ |
| 1136 | ((__mmask8)__builtin_ia32_fpclassps128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 1137 | (__mmask8)(U))) |
| 1138 | |
| 1139 | #define _mm_fpclass_ps_mask(A, imm) \ |
| 1140 | ((__mmask8)__builtin_ia32_fpclassps128_mask((__v4sf)(__m128)(A), (int)(imm), \ |
| 1141 | (__mmask8)-1)) |
| 1142 | |
| 1143 | #define _mm256_mask_fpclass_ps_mask(U, A, imm) \ |
| 1144 | ((__mmask8)__builtin_ia32_fpclassps256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 1145 | (__mmask8)(U))) |
| 1146 | |
| 1147 | #define _mm256_fpclass_ps_mask(A, imm) \ |
| 1148 | ((__mmask8)__builtin_ia32_fpclassps256_mask((__v8sf)(__m256)(A), (int)(imm), \ |
| 1149 | (__mmask8)-1)) |
| 1150 | |
| 1151 | #undef __DEFAULT_FN_ATTRS128 |
| 1152 | #undef __DEFAULT_FN_ATTRS256 |
| 1153 | #undef __DEFAULT_FN_ATTRS256_CONSTEXPR |
| 1154 | #undef __DEFAULT_FN_ATTRS128_CONSTEXPR |
| 1155 | |
| 1156 | #endif |