1 /*===---- avxintrin.h - AVX intrinsics -------------------------------------=== 2 * 3 * Permission is hereby granted, free of charge, to any person obtaining a copy 4 * of this software and associated documentation files (the "Software"), to deal 5 * in the Software without restriction, including without limitation the rights 6 * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 7 * copies of the Software, and to permit persons to whom the Software is 8 * furnished to do so, subject to the following conditions: 9 * 10 * The above copyright notice and this permission notice shall be included in 11 * all copies or substantial portions of the Software. 12 * 13 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 14 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 15 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 16 * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 17 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 18 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN 19 * THE SOFTWARE. 20 * 21 *===-----------------------------------------------------------------------=== 22 */ 23 24 #ifndef __IMMINTRIN_H 25 #error "Never use <avxintrin.h> directly; include <immintrin.h> instead." 26 #endif 27 28 #ifndef __AVXINTRIN_H 29 #define __AVXINTRIN_H 30 31 typedef double __v4df __attribute__ ((__vector_size__ (32))); 32 typedef float __v8sf __attribute__ ((__vector_size__ (32))); 33 typedef long long __v4di __attribute__ ((__vector_size__ (32))); 34 typedef int __v8si __attribute__ ((__vector_size__ (32))); 35 typedef short __v16hi __attribute__ ((__vector_size__ (32))); 36 typedef char __v32qi __attribute__ ((__vector_size__ (32))); 37 38 typedef float __m256 __attribute__ ((__vector_size__ (32))); 39 typedef double __m256d __attribute__((__vector_size__(32))); 40 typedef long long __m256i __attribute__((__vector_size__(32))); 41 42 /* Arithmetic */ 43 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 44 _mm256_add_pd(__m256d __a, __m256d __b) 45 { 46 return __a+__b; 47 } 48 49 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 50 _mm256_add_ps(__m256 __a, __m256 __b) 51 { 52 return __a+__b; 53 } 54 55 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 56 _mm256_sub_pd(__m256d __a, __m256d __b) 57 { 58 return __a-__b; 59 } 60 61 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 62 _mm256_sub_ps(__m256 __a, __m256 __b) 63 { 64 return __a-__b; 65 } 66 67 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 68 _mm256_addsub_pd(__m256d __a, __m256d __b) 69 { 70 return (__m256d)__builtin_ia32_addsubpd256((__v4df)__a, (__v4df)__b); 71 } 72 73 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 74 _mm256_addsub_ps(__m256 __a, __m256 __b) 75 { 76 return (__m256)__builtin_ia32_addsubps256((__v8sf)__a, (__v8sf)__b); 77 } 78 79 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 80 _mm256_div_pd(__m256d __a, __m256d __b) 81 { 82 return __a / __b; 83 } 84 85 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 86 _mm256_div_ps(__m256 __a, __m256 __b) 87 { 88 return __a / __b; 89 } 90 91 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 92 _mm256_max_pd(__m256d __a, __m256d __b) 93 { 94 return (__m256d)__builtin_ia32_maxpd256((__v4df)__a, (__v4df)__b); 95 } 96 97 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 98 _mm256_max_ps(__m256 __a, __m256 __b) 99 { 100 return (__m256)__builtin_ia32_maxps256((__v8sf)__a, (__v8sf)__b); 101 } 102 103 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 104 _mm256_min_pd(__m256d __a, __m256d __b) 105 { 106 return (__m256d)__builtin_ia32_minpd256((__v4df)__a, (__v4df)__b); 107 } 108 109 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 110 _mm256_min_ps(__m256 __a, __m256 __b) 111 { 112 return (__m256)__builtin_ia32_minps256((__v8sf)__a, (__v8sf)__b); 113 } 114 115 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 116 _mm256_mul_pd(__m256d __a, __m256d __b) 117 { 118 return __a * __b; 119 } 120 121 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 122 _mm256_mul_ps(__m256 __a, __m256 __b) 123 { 124 return __a * __b; 125 } 126 127 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 128 _mm256_sqrt_pd(__m256d __a) 129 { 130 return (__m256d)__builtin_ia32_sqrtpd256((__v4df)__a); 131 } 132 133 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 134 _mm256_sqrt_ps(__m256 __a) 135 { 136 return (__m256)__builtin_ia32_sqrtps256((__v8sf)__a); 137 } 138 139 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 140 _mm256_rsqrt_ps(__m256 __a) 141 { 142 return (__m256)__builtin_ia32_rsqrtps256((__v8sf)__a); 143 } 144 145 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 146 _mm256_rcp_ps(__m256 __a) 147 { 148 return (__m256)__builtin_ia32_rcpps256((__v8sf)__a); 149 } 150 151 #define _mm256_round_pd(V, M) __extension__ ({ \ 152 __m256d __V = (V); \ 153 (__m256d)__builtin_ia32_roundpd256((__v4df)__V, (M)); }) 154 155 #define _mm256_round_ps(V, M) __extension__ ({ \ 156 __m256 __V = (V); \ 157 (__m256)__builtin_ia32_roundps256((__v8sf)__V, (M)); }) 158 159 #define _mm256_ceil_pd(V) _mm256_round_pd((V), _MM_FROUND_CEIL) 160 #define _mm256_floor_pd(V) _mm256_round_pd((V), _MM_FROUND_FLOOR) 161 #define _mm256_ceil_ps(V) _mm256_round_ps((V), _MM_FROUND_CEIL) 162 #define _mm256_floor_ps(V) _mm256_round_ps((V), _MM_FROUND_FLOOR) 163 164 /* Logical */ 165 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 166 _mm256_and_pd(__m256d __a, __m256d __b) 167 { 168 return (__m256d)((__v4di)__a & (__v4di)__b); 169 } 170 171 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 172 _mm256_and_ps(__m256 __a, __m256 __b) 173 { 174 return (__m256)((__v8si)__a & (__v8si)__b); 175 } 176 177 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 178 _mm256_andnot_pd(__m256d __a, __m256d __b) 179 { 180 return (__m256d)(~(__v4di)__a & (__v4di)__b); 181 } 182 183 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 184 _mm256_andnot_ps(__m256 __a, __m256 __b) 185 { 186 return (__m256)(~(__v8si)__a & (__v8si)__b); 187 } 188 189 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 190 _mm256_or_pd(__m256d __a, __m256d __b) 191 { 192 return (__m256d)((__v4di)__a | (__v4di)__b); 193 } 194 195 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 196 _mm256_or_ps(__m256 __a, __m256 __b) 197 { 198 return (__m256)((__v8si)__a | (__v8si)__b); 199 } 200 201 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 202 _mm256_xor_pd(__m256d __a, __m256d __b) 203 { 204 return (__m256d)((__v4di)__a ^ (__v4di)__b); 205 } 206 207 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 208 _mm256_xor_ps(__m256 __a, __m256 __b) 209 { 210 return (__m256)((__v8si)__a ^ (__v8si)__b); 211 } 212 213 /* Horizontal arithmetic */ 214 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 215 _mm256_hadd_pd(__m256d __a, __m256d __b) 216 { 217 return (__m256d)__builtin_ia32_haddpd256((__v4df)__a, (__v4df)__b); 218 } 219 220 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 221 _mm256_hadd_ps(__m256 __a, __m256 __b) 222 { 223 return (__m256)__builtin_ia32_haddps256((__v8sf)__a, (__v8sf)__b); 224 } 225 226 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 227 _mm256_hsub_pd(__m256d __a, __m256d __b) 228 { 229 return (__m256d)__builtin_ia32_hsubpd256((__v4df)__a, (__v4df)__b); 230 } 231 232 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 233 _mm256_hsub_ps(__m256 __a, __m256 __b) 234 { 235 return (__m256)__builtin_ia32_hsubps256((__v8sf)__a, (__v8sf)__b); 236 } 237 238 /* Vector permutations */ 239 static __inline __m128d __attribute__((__always_inline__, __nodebug__)) 240 _mm_permutevar_pd(__m128d __a, __m128i __c) 241 { 242 return (__m128d)__builtin_ia32_vpermilvarpd((__v2df)__a, (__v2di)__c); 243 } 244 245 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 246 _mm256_permutevar_pd(__m256d __a, __m256i __c) 247 { 248 return (__m256d)__builtin_ia32_vpermilvarpd256((__v4df)__a, (__v4di)__c); 249 } 250 251 static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 252 _mm_permutevar_ps(__m128 __a, __m128i __c) 253 { 254 return (__m128)__builtin_ia32_vpermilvarps((__v4sf)__a, (__v4si)__c); 255 } 256 257 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 258 _mm256_permutevar_ps(__m256 __a, __m256i __c) 259 { 260 return (__m256)__builtin_ia32_vpermilvarps256((__v8sf)__a, 261 (__v8si)__c); 262 } 263 264 #define _mm_permute_pd(A, C) __extension__ ({ \ 265 __m128d __A = (A); \ 266 (__m128d)__builtin_shufflevector((__v2df)__A, (__v2df) _mm_setzero_pd(), \ 267 (C) & 0x1, ((C) & 0x2) >> 1); }) 268 269 #define _mm256_permute_pd(A, C) __extension__ ({ \ 270 __m256d __A = (A); \ 271 (__m256d)__builtin_shufflevector((__v4df)__A, (__v4df) _mm256_setzero_pd(), \ 272 (C) & 0x1, ((C) & 0x2) >> 1, \ 273 2 + (((C) & 0x4) >> 2), \ 274 2 + (((C) & 0x8) >> 3)); }) 275 276 #define _mm_permute_ps(A, C) __extension__ ({ \ 277 __m128 __A = (A); \ 278 (__m128)__builtin_shufflevector((__v4sf)__A, (__v4sf) _mm_setzero_ps(), \ 279 (C) & 0x3, ((C) & 0xc) >> 2, \ 280 ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6); }) 281 282 #define _mm256_permute_ps(A, C) __extension__ ({ \ 283 __m256 __A = (A); \ 284 (__m256)__builtin_shufflevector((__v8sf)__A, (__v8sf) _mm256_setzero_ps(), \ 285 (C) & 0x3, ((C) & 0xc) >> 2, \ 286 ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6, \ 287 4 + (((C) & 0x03) >> 0), \ 288 4 + (((C) & 0x0c) >> 2), \ 289 4 + (((C) & 0x30) >> 4), \ 290 4 + (((C) & 0xc0) >> 6)); }) 291 292 #define _mm256_permute2f128_pd(V1, V2, M) __extension__ ({ \ 293 __m256d __V1 = (V1); \ 294 __m256d __V2 = (V2); \ 295 (__m256d)__builtin_ia32_vperm2f128_pd256((__v4df)__V1, (__v4df)__V2, (M)); }) 296 297 #define _mm256_permute2f128_ps(V1, V2, M) __extension__ ({ \ 298 __m256 __V1 = (V1); \ 299 __m256 __V2 = (V2); \ 300 (__m256)__builtin_ia32_vperm2f128_ps256((__v8sf)__V1, (__v8sf)__V2, (M)); }) 301 302 #define _mm256_permute2f128_si256(V1, V2, M) __extension__ ({ \ 303 __m256i __V1 = (V1); \ 304 __m256i __V2 = (V2); \ 305 (__m256i)__builtin_ia32_vperm2f128_si256((__v8si)__V1, (__v8si)__V2, (M)); }) 306 307 /* Vector Blend */ 308 #define _mm256_blend_pd(V1, V2, M) __extension__ ({ \ 309 __m256d __V1 = (V1); \ 310 __m256d __V2 = (V2); \ 311 (__m256d)__builtin_shufflevector((__v4df)__V1, (__v4df)__V2, \ 312 (((M) & 0x01) ? 4 : 0), \ 313 (((M) & 0x02) ? 5 : 1), \ 314 (((M) & 0x04) ? 6 : 2), \ 315 (((M) & 0x08) ? 7 : 3)); }) 316 317 #define _mm256_blend_ps(V1, V2, M) __extension__ ({ \ 318 __m256 __V1 = (V1); \ 319 __m256 __V2 = (V2); \ 320 (__m256)__builtin_shufflevector((__v8sf)__V1, (__v8sf)__V2, \ 321 (((M) & 0x01) ? 8 : 0), \ 322 (((M) & 0x02) ? 9 : 1), \ 323 (((M) & 0x04) ? 10 : 2), \ 324 (((M) & 0x08) ? 11 : 3), \ 325 (((M) & 0x10) ? 12 : 4), \ 326 (((M) & 0x20) ? 13 : 5), \ 327 (((M) & 0x40) ? 14 : 6), \ 328 (((M) & 0x80) ? 15 : 7)); }) 329 330 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 331 _mm256_blendv_pd(__m256d __a, __m256d __b, __m256d __c) 332 { 333 return (__m256d)__builtin_ia32_blendvpd256( 334 (__v4df)__a, (__v4df)__b, (__v4df)__c); 335 } 336 337 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 338 _mm256_blendv_ps(__m256 __a, __m256 __b, __m256 __c) 339 { 340 return (__m256)__builtin_ia32_blendvps256( 341 (__v8sf)__a, (__v8sf)__b, (__v8sf)__c); 342 } 343 344 /* Vector Dot Product */ 345 #define _mm256_dp_ps(V1, V2, M) __extension__ ({ \ 346 __m256 __V1 = (V1); \ 347 __m256 __V2 = (V2); \ 348 (__m256)__builtin_ia32_dpps256((__v8sf)__V1, (__v8sf)__V2, (M)); }) 349 350 /* Vector shuffle */ 351 #define _mm256_shuffle_ps(a, b, mask) __extension__ ({ \ 352 __m256 __a = (a); \ 353 __m256 __b = (b); \ 354 (__m256)__builtin_shufflevector((__v8sf)__a, (__v8sf)__b, \ 355 (mask) & 0x3, ((mask) & 0xc) >> 2, \ 356 (((mask) & 0x30) >> 4) + 8, (((mask) & 0xc0) >> 6) + 8, \ 357 ((mask) & 0x3) + 4, (((mask) & 0xc) >> 2) + 4, \ 358 (((mask) & 0x30) >> 4) + 12, (((mask) & 0xc0) >> 6) + 12); }) 359 360 #define _mm256_shuffle_pd(a, b, mask) __extension__ ({ \ 361 __m256d __a = (a); \ 362 __m256d __b = (b); \ 363 (__m256d)__builtin_shufflevector((__v4df)__a, (__v4df)__b, \ 364 (mask) & 0x1, \ 365 (((mask) & 0x2) >> 1) + 4, \ 366 (((mask) & 0x4) >> 2) + 2, \ 367 (((mask) & 0x8) >> 3) + 6); }) 368 369 /* Compare */ 370 #define _CMP_EQ_OQ 0x00 /* Equal (ordered, non-signaling) */ 371 #define _CMP_LT_OS 0x01 /* Less-than (ordered, signaling) */ 372 #define _CMP_LE_OS 0x02 /* Less-than-or-equal (ordered, signaling) */ 373 #define _CMP_UNORD_Q 0x03 /* Unordered (non-signaling) */ 374 #define _CMP_NEQ_UQ 0x04 /* Not-equal (unordered, non-signaling) */ 375 #define _CMP_NLT_US 0x05 /* Not-less-than (unordered, signaling) */ 376 #define _CMP_NLE_US 0x06 /* Not-less-than-or-equal (unordered, signaling) */ 377 #define _CMP_ORD_Q 0x07 /* Ordered (nonsignaling) */ 378 #define _CMP_EQ_UQ 0x08 /* Equal (unordered, non-signaling) */ 379 #define _CMP_NGE_US 0x09 /* Not-greater-than-or-equal (unord, signaling) */ 380 #define _CMP_NGT_US 0x0a /* Not-greater-than (unordered, signaling) */ 381 #define _CMP_FALSE_OQ 0x0b /* False (ordered, non-signaling) */ 382 #define _CMP_NEQ_OQ 0x0c /* Not-equal (ordered, non-signaling) */ 383 #define _CMP_GE_OS 0x0d /* Greater-than-or-equal (ordered, signaling) */ 384 #define _CMP_GT_OS 0x0e /* Greater-than (ordered, signaling) */ 385 #define _CMP_TRUE_UQ 0x0f /* True (unordered, non-signaling) */ 386 #define _CMP_EQ_OS 0x10 /* Equal (ordered, signaling) */ 387 #define _CMP_LT_OQ 0x11 /* Less-than (ordered, non-signaling) */ 388 #define _CMP_LE_OQ 0x12 /* Less-than-or-equal (ordered, non-signaling) */ 389 #define _CMP_UNORD_S 0x13 /* Unordered (signaling) */ 390 #define _CMP_NEQ_US 0x14 /* Not-equal (unordered, signaling) */ 391 #define _CMP_NLT_UQ 0x15 /* Not-less-than (unordered, non-signaling) */ 392 #define _CMP_NLE_UQ 0x16 /* Not-less-than-or-equal (unord, non-signaling) */ 393 #define _CMP_ORD_S 0x17 /* Ordered (signaling) */ 394 #define _CMP_EQ_US 0x18 /* Equal (unordered, signaling) */ 395 #define _CMP_NGE_UQ 0x19 /* Not-greater-than-or-equal (unord, non-sign) */ 396 #define _CMP_NGT_UQ 0x1a /* Not-greater-than (unordered, non-signaling) */ 397 #define _CMP_FALSE_OS 0x1b /* False (ordered, signaling) */ 398 #define _CMP_NEQ_OS 0x1c /* Not-equal (ordered, signaling) */ 399 #define _CMP_GE_OQ 0x1d /* Greater-than-or-equal (ordered, non-signaling) */ 400 #define _CMP_GT_OQ 0x1e /* Greater-than (ordered, non-signaling) */ 401 #define _CMP_TRUE_US 0x1f /* True (unordered, signaling) */ 402 403 #define _mm_cmp_pd(a, b, c) __extension__ ({ \ 404 __m128d __a = (a); \ 405 __m128d __b = (b); \ 406 (__m128d)__builtin_ia32_cmppd((__v2df)__a, (__v2df)__b, (c)); }) 407 408 #define _mm_cmp_ps(a, b, c) __extension__ ({ \ 409 __m128 __a = (a); \ 410 __m128 __b = (b); \ 411 (__m128)__builtin_ia32_cmpps((__v4sf)__a, (__v4sf)__b, (c)); }) 412 413 #define _mm256_cmp_pd(a, b, c) __extension__ ({ \ 414 __m256d __a = (a); \ 415 __m256d __b = (b); \ 416 (__m256d)__builtin_ia32_cmppd256((__v4df)__a, (__v4df)__b, (c)); }) 417 418 #define _mm256_cmp_ps(a, b, c) __extension__ ({ \ 419 __m256 __a = (a); \ 420 __m256 __b = (b); \ 421 (__m256)__builtin_ia32_cmpps256((__v8sf)__a, (__v8sf)__b, (c)); }) 422 423 #define _mm_cmp_sd(a, b, c) __extension__ ({ \ 424 __m128d __a = (a); \ 425 __m128d __b = (b); \ 426 (__m128d)__builtin_ia32_cmpsd((__v2df)__a, (__v2df)__b, (c)); }) 427 428 #define _mm_cmp_ss(a, b, c) __extension__ ({ \ 429 __m128 __a = (a); \ 430 __m128 __b = (b); \ 431 (__m128)__builtin_ia32_cmpss((__v4sf)__a, (__v4sf)__b, (c)); }) 432 433 /* Vector extract */ 434 #define _mm256_extractf128_pd(A, O) __extension__ ({ \ 435 __m256d __A = (A); \ 436 (__m128d)__builtin_ia32_vextractf128_pd256((__v4df)__A, (O)); }) 437 438 #define _mm256_extractf128_ps(A, O) __extension__ ({ \ 439 __m256 __A = (A); \ 440 (__m128)__builtin_ia32_vextractf128_ps256((__v8sf)__A, (O)); }) 441 442 #define _mm256_extractf128_si256(A, O) __extension__ ({ \ 443 __m256i __A = (A); \ 444 (__m128i)__builtin_ia32_vextractf128_si256((__v8si)__A, (O)); }) 445 446 static __inline int __attribute__((__always_inline__, __nodebug__)) 447 _mm256_extract_epi32(__m256i __a, int const __imm) 448 { 449 __v8si __b = (__v8si)__a; 450 return __b[__imm & 7]; 451 } 452 453 static __inline int __attribute__((__always_inline__, __nodebug__)) 454 _mm256_extract_epi16(__m256i __a, int const __imm) 455 { 456 __v16hi __b = (__v16hi)__a; 457 return __b[__imm & 15]; 458 } 459 460 static __inline int __attribute__((__always_inline__, __nodebug__)) 461 _mm256_extract_epi8(__m256i __a, int const __imm) 462 { 463 __v32qi __b = (__v32qi)__a; 464 return __b[__imm & 31]; 465 } 466 467 #ifdef __x86_64__ 468 static __inline long long __attribute__((__always_inline__, __nodebug__)) 469 _mm256_extract_epi64(__m256i __a, const int __imm) 470 { 471 __v4di __b = (__v4di)__a; 472 return __b[__imm & 3]; 473 } 474 #endif 475 476 /* Vector insert */ 477 #define _mm256_insertf128_pd(V1, V2, O) __extension__ ({ \ 478 __m256d __V1 = (V1); \ 479 __m128d __V2 = (V2); \ 480 (__m256d)__builtin_ia32_vinsertf128_pd256((__v4df)__V1, (__v2df)__V2, (O)); }) 481 482 #define _mm256_insertf128_ps(V1, V2, O) __extension__ ({ \ 483 __m256 __V1 = (V1); \ 484 __m128 __V2 = (V2); \ 485 (__m256)__builtin_ia32_vinsertf128_ps256((__v8sf)__V1, (__v4sf)__V2, (O)); }) 486 487 #define _mm256_insertf128_si256(V1, V2, O) __extension__ ({ \ 488 __m256i __V1 = (V1); \ 489 __m128i __V2 = (V2); \ 490 (__m256i)__builtin_ia32_vinsertf128_si256((__v8si)__V1, (__v4si)__V2, (O)); }) 491 492 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 493 _mm256_insert_epi32(__m256i __a, int __b, int const __imm) 494 { 495 __v8si __c = (__v8si)__a; 496 __c[__imm & 7] = __b; 497 return (__m256i)__c; 498 } 499 500 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 501 _mm256_insert_epi16(__m256i __a, int __b, int const __imm) 502 { 503 __v16hi __c = (__v16hi)__a; 504 __c[__imm & 15] = __b; 505 return (__m256i)__c; 506 } 507 508 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 509 _mm256_insert_epi8(__m256i __a, int __b, int const __imm) 510 { 511 __v32qi __c = (__v32qi)__a; 512 __c[__imm & 31] = __b; 513 return (__m256i)__c; 514 } 515 516 #ifdef __x86_64__ 517 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 518 _mm256_insert_epi64(__m256i __a, int __b, int const __imm) 519 { 520 __v4di __c = (__v4di)__a; 521 __c[__imm & 3] = __b; 522 return (__m256i)__c; 523 } 524 #endif 525 526 /* Conversion */ 527 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 528 _mm256_cvtepi32_pd(__m128i __a) 529 { 530 return (__m256d)__builtin_ia32_cvtdq2pd256((__v4si) __a); 531 } 532 533 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 534 _mm256_cvtepi32_ps(__m256i __a) 535 { 536 return (__m256)__builtin_ia32_cvtdq2ps256((__v8si) __a); 537 } 538 539 static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 540 _mm256_cvtpd_ps(__m256d __a) 541 { 542 return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) __a); 543 } 544 545 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 546 _mm256_cvtps_epi32(__m256 __a) 547 { 548 return (__m256i)__builtin_ia32_cvtps2dq256((__v8sf) __a); 549 } 550 551 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 552 _mm256_cvtps_pd(__m128 __a) 553 { 554 return (__m256d)__builtin_ia32_cvtps2pd256((__v4sf) __a); 555 } 556 557 static __inline __m128i __attribute__((__always_inline__, __nodebug__)) 558 _mm256_cvttpd_epi32(__m256d __a) 559 { 560 return (__m128i)__builtin_ia32_cvttpd2dq256((__v4df) __a); 561 } 562 563 static __inline __m128i __attribute__((__always_inline__, __nodebug__)) 564 _mm256_cvtpd_epi32(__m256d __a) 565 { 566 return (__m128i)__builtin_ia32_cvtpd2dq256((__v4df) __a); 567 } 568 569 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 570 _mm256_cvttps_epi32(__m256 __a) 571 { 572 return (__m256i)__builtin_ia32_cvttps2dq256((__v8sf) __a); 573 } 574 575 /* Vector replicate */ 576 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 577 _mm256_movehdup_ps(__m256 __a) 578 { 579 return __builtin_shufflevector(__a, __a, 1, 1, 3, 3, 5, 5, 7, 7); 580 } 581 582 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 583 _mm256_moveldup_ps(__m256 __a) 584 { 585 return __builtin_shufflevector(__a, __a, 0, 0, 2, 2, 4, 4, 6, 6); 586 } 587 588 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 589 _mm256_movedup_pd(__m256d __a) 590 { 591 return __builtin_shufflevector(__a, __a, 0, 0, 2, 2); 592 } 593 594 /* Unpack and Interleave */ 595 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 596 _mm256_unpackhi_pd(__m256d __a, __m256d __b) 597 { 598 return __builtin_shufflevector(__a, __b, 1, 5, 1+2, 5+2); 599 } 600 601 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 602 _mm256_unpacklo_pd(__m256d __a, __m256d __b) 603 { 604 return __builtin_shufflevector(__a, __b, 0, 4, 0+2, 4+2); 605 } 606 607 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 608 _mm256_unpackhi_ps(__m256 __a, __m256 __b) 609 { 610 return __builtin_shufflevector(__a, __b, 2, 10, 2+1, 10+1, 6, 14, 6+1, 14+1); 611 } 612 613 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 614 _mm256_unpacklo_ps(__m256 __a, __m256 __b) 615 { 616 return __builtin_shufflevector(__a, __b, 0, 8, 0+1, 8+1, 4, 12, 4+1, 12+1); 617 } 618 619 /* Bit Test */ 620 static __inline int __attribute__((__always_inline__, __nodebug__)) 621 _mm_testz_pd(__m128d __a, __m128d __b) 622 { 623 return __builtin_ia32_vtestzpd((__v2df)__a, (__v2df)__b); 624 } 625 626 static __inline int __attribute__((__always_inline__, __nodebug__)) 627 _mm_testc_pd(__m128d __a, __m128d __b) 628 { 629 return __builtin_ia32_vtestcpd((__v2df)__a, (__v2df)__b); 630 } 631 632 static __inline int __attribute__((__always_inline__, __nodebug__)) 633 _mm_testnzc_pd(__m128d __a, __m128d __b) 634 { 635 return __builtin_ia32_vtestnzcpd((__v2df)__a, (__v2df)__b); 636 } 637 638 static __inline int __attribute__((__always_inline__, __nodebug__)) 639 _mm_testz_ps(__m128 __a, __m128 __b) 640 { 641 return __builtin_ia32_vtestzps((__v4sf)__a, (__v4sf)__b); 642 } 643 644 static __inline int __attribute__((__always_inline__, __nodebug__)) 645 _mm_testc_ps(__m128 __a, __m128 __b) 646 { 647 return __builtin_ia32_vtestcps((__v4sf)__a, (__v4sf)__b); 648 } 649 650 static __inline int __attribute__((__always_inline__, __nodebug__)) 651 _mm_testnzc_ps(__m128 __a, __m128 __b) 652 { 653 return __builtin_ia32_vtestnzcps((__v4sf)__a, (__v4sf)__b); 654 } 655 656 static __inline int __attribute__((__always_inline__, __nodebug__)) 657 _mm256_testz_pd(__m256d __a, __m256d __b) 658 { 659 return __builtin_ia32_vtestzpd256((__v4df)__a, (__v4df)__b); 660 } 661 662 static __inline int __attribute__((__always_inline__, __nodebug__)) 663 _mm256_testc_pd(__m256d __a, __m256d __b) 664 { 665 return __builtin_ia32_vtestcpd256((__v4df)__a, (__v4df)__b); 666 } 667 668 static __inline int __attribute__((__always_inline__, __nodebug__)) 669 _mm256_testnzc_pd(__m256d __a, __m256d __b) 670 { 671 return __builtin_ia32_vtestnzcpd256((__v4df)__a, (__v4df)__b); 672 } 673 674 static __inline int __attribute__((__always_inline__, __nodebug__)) 675 _mm256_testz_ps(__m256 __a, __m256 __b) 676 { 677 return __builtin_ia32_vtestzps256((__v8sf)__a, (__v8sf)__b); 678 } 679 680 static __inline int __attribute__((__always_inline__, __nodebug__)) 681 _mm256_testc_ps(__m256 __a, __m256 __b) 682 { 683 return __builtin_ia32_vtestcps256((__v8sf)__a, (__v8sf)__b); 684 } 685 686 static __inline int __attribute__((__always_inline__, __nodebug__)) 687 _mm256_testnzc_ps(__m256 __a, __m256 __b) 688 { 689 return __builtin_ia32_vtestnzcps256((__v8sf)__a, (__v8sf)__b); 690 } 691 692 static __inline int __attribute__((__always_inline__, __nodebug__)) 693 _mm256_testz_si256(__m256i __a, __m256i __b) 694 { 695 return __builtin_ia32_ptestz256((__v4di)__a, (__v4di)__b); 696 } 697 698 static __inline int __attribute__((__always_inline__, __nodebug__)) 699 _mm256_testc_si256(__m256i __a, __m256i __b) 700 { 701 return __builtin_ia32_ptestc256((__v4di)__a, (__v4di)__b); 702 } 703 704 static __inline int __attribute__((__always_inline__, __nodebug__)) 705 _mm256_testnzc_si256(__m256i __a, __m256i __b) 706 { 707 return __builtin_ia32_ptestnzc256((__v4di)__a, (__v4di)__b); 708 } 709 710 /* Vector extract sign mask */ 711 static __inline int __attribute__((__always_inline__, __nodebug__)) 712 _mm256_movemask_pd(__m256d __a) 713 { 714 return __builtin_ia32_movmskpd256((__v4df)__a); 715 } 716 717 static __inline int __attribute__((__always_inline__, __nodebug__)) 718 _mm256_movemask_ps(__m256 __a) 719 { 720 return __builtin_ia32_movmskps256((__v8sf)__a); 721 } 722 723 /* Vector __zero */ 724 static __inline void __attribute__((__always_inline__, __nodebug__)) 725 _mm256_zeroall(void) 726 { 727 __builtin_ia32_vzeroall(); 728 } 729 730 static __inline void __attribute__((__always_inline__, __nodebug__)) 731 _mm256_zeroupper(void) 732 { 733 __builtin_ia32_vzeroupper(); 734 } 735 736 /* Vector load with broadcast */ 737 static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 738 _mm_broadcast_ss(float const *__a) 739 { 740 float __f = *__a; 741 return (__m128)(__v4sf){ __f, __f, __f, __f }; 742 } 743 744 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 745 _mm256_broadcast_sd(double const *__a) 746 { 747 double __d = *__a; 748 return (__m256d)(__v4df){ __d, __d, __d, __d }; 749 } 750 751 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 752 _mm256_broadcast_ss(float const *__a) 753 { 754 float __f = *__a; 755 return (__m256)(__v8sf){ __f, __f, __f, __f, __f, __f, __f, __f }; 756 } 757 758 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 759 _mm256_broadcast_pd(__m128d const *__a) 760 { 761 return (__m256d)__builtin_ia32_vbroadcastf128_pd256(__a); 762 } 763 764 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 765 _mm256_broadcast_ps(__m128 const *__a) 766 { 767 return (__m256)__builtin_ia32_vbroadcastf128_ps256(__a); 768 } 769 770 /* SIMD load ops */ 771 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 772 _mm256_load_pd(double const *__p) 773 { 774 return *(__m256d *)__p; 775 } 776 777 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 778 _mm256_load_ps(float const *__p) 779 { 780 return *(__m256 *)__p; 781 } 782 783 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 784 _mm256_loadu_pd(double const *__p) 785 { 786 struct __loadu_pd { 787 __m256d __v; 788 } __attribute__((packed, may_alias)); 789 return ((struct __loadu_pd*)__p)->__v; 790 } 791 792 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 793 _mm256_loadu_ps(float const *__p) 794 { 795 struct __loadu_ps { 796 __m256 __v; 797 } __attribute__((packed, may_alias)); 798 return ((struct __loadu_ps*)__p)->__v; 799 } 800 801 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 802 _mm256_load_si256(__m256i const *__p) 803 { 804 return *__p; 805 } 806 807 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 808 _mm256_loadu_si256(__m256i const *__p) 809 { 810 struct __loadu_si256 { 811 __m256i __v; 812 } __attribute__((packed, may_alias)); 813 return ((struct __loadu_si256*)__p)->__v; 814 } 815 816 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 817 _mm256_lddqu_si256(__m256i const *__p) 818 { 819 return (__m256i)__builtin_ia32_lddqu256((char const *)__p); 820 } 821 822 /* SIMD store ops */ 823 static __inline void __attribute__((__always_inline__, __nodebug__)) 824 _mm256_store_pd(double *__p, __m256d __a) 825 { 826 *(__m256d *)__p = __a; 827 } 828 829 static __inline void __attribute__((__always_inline__, __nodebug__)) 830 _mm256_store_ps(float *__p, __m256 __a) 831 { 832 *(__m256 *)__p = __a; 833 } 834 835 static __inline void __attribute__((__always_inline__, __nodebug__)) 836 _mm256_storeu_pd(double *__p, __m256d __a) 837 { 838 __builtin_ia32_storeupd256(__p, (__v4df)__a); 839 } 840 841 static __inline void __attribute__((__always_inline__, __nodebug__)) 842 _mm256_storeu_ps(float *__p, __m256 __a) 843 { 844 __builtin_ia32_storeups256(__p, (__v8sf)__a); 845 } 846 847 static __inline void __attribute__((__always_inline__, __nodebug__)) 848 _mm256_store_si256(__m256i *__p, __m256i __a) 849 { 850 *__p = __a; 851 } 852 853 static __inline void __attribute__((__always_inline__, __nodebug__)) 854 _mm256_storeu_si256(__m256i *__p, __m256i __a) 855 { 856 __builtin_ia32_storedqu256((char *)__p, (__v32qi)__a); 857 } 858 859 /* Conditional load ops */ 860 static __inline __m128d __attribute__((__always_inline__, __nodebug__)) 861 _mm_maskload_pd(double const *__p, __m128d __m) 862 { 863 return (__m128d)__builtin_ia32_maskloadpd((const __v2df *)__p, (__v2df)__m); 864 } 865 866 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 867 _mm256_maskload_pd(double const *__p, __m256d __m) 868 { 869 return (__m256d)__builtin_ia32_maskloadpd256((const __v4df *)__p, 870 (__v4df)__m); 871 } 872 873 static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 874 _mm_maskload_ps(float const *__p, __m128 __m) 875 { 876 return (__m128)__builtin_ia32_maskloadps((const __v4sf *)__p, (__v4sf)__m); 877 } 878 879 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 880 _mm256_maskload_ps(float const *__p, __m256 __m) 881 { 882 return (__m256)__builtin_ia32_maskloadps256((const __v8sf *)__p, (__v8sf)__m); 883 } 884 885 /* Conditional store ops */ 886 static __inline void __attribute__((__always_inline__, __nodebug__)) 887 _mm256_maskstore_ps(float *__p, __m256 __m, __m256 __a) 888 { 889 __builtin_ia32_maskstoreps256((__v8sf *)__p, (__v8sf)__m, (__v8sf)__a); 890 } 891 892 static __inline void __attribute__((__always_inline__, __nodebug__)) 893 _mm_maskstore_pd(double *__p, __m128d __m, __m128d __a) 894 { 895 __builtin_ia32_maskstorepd((__v2df *)__p, (__v2df)__m, (__v2df)__a); 896 } 897 898 static __inline void __attribute__((__always_inline__, __nodebug__)) 899 _mm256_maskstore_pd(double *__p, __m256d __m, __m256d __a) 900 { 901 __builtin_ia32_maskstorepd256((__v4df *)__p, (__v4df)__m, (__v4df)__a); 902 } 903 904 static __inline void __attribute__((__always_inline__, __nodebug__)) 905 _mm_maskstore_ps(float *__p, __m128 __m, __m128 __a) 906 { 907 __builtin_ia32_maskstoreps((__v4sf *)__p, (__v4sf)__m, (__v4sf)__a); 908 } 909 910 /* Cacheability support ops */ 911 static __inline void __attribute__((__always_inline__, __nodebug__)) 912 _mm256_stream_si256(__m256i *__a, __m256i __b) 913 { 914 __builtin_ia32_movntdq256((__v4di *)__a, (__v4di)__b); 915 } 916 917 static __inline void __attribute__((__always_inline__, __nodebug__)) 918 _mm256_stream_pd(double *__a, __m256d __b) 919 { 920 __builtin_ia32_movntpd256(__a, (__v4df)__b); 921 } 922 923 static __inline void __attribute__((__always_inline__, __nodebug__)) 924 _mm256_stream_ps(float *__p, __m256 __a) 925 { 926 __builtin_ia32_movntps256(__p, (__v8sf)__a); 927 } 928 929 /* Create vectors */ 930 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 931 _mm256_set_pd(double __a, double __b, double __c, double __d) 932 { 933 return (__m256d){ __d, __c, __b, __a }; 934 } 935 936 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 937 _mm256_set_ps(float __a, float __b, float __c, float __d, 938 float __e, float __f, float __g, float __h) 939 { 940 return (__m256){ __h, __g, __f, __e, __d, __c, __b, __a }; 941 } 942 943 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 944 _mm256_set_epi32(int __i0, int __i1, int __i2, int __i3, 945 int __i4, int __i5, int __i6, int __i7) 946 { 947 return (__m256i)(__v8si){ __i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0 }; 948 } 949 950 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 951 _mm256_set_epi16(short __w15, short __w14, short __w13, short __w12, 952 short __w11, short __w10, short __w09, short __w08, 953 short __w07, short __w06, short __w05, short __w04, 954 short __w03, short __w02, short __w01, short __w00) 955 { 956 return (__m256i)(__v16hi){ __w00, __w01, __w02, __w03, __w04, __w05, __w06, 957 __w07, __w08, __w09, __w10, __w11, __w12, __w13, __w14, __w15 }; 958 } 959 960 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 961 _mm256_set_epi8(char __b31, char __b30, char __b29, char __b28, 962 char __b27, char __b26, char __b25, char __b24, 963 char __b23, char __b22, char __b21, char __b20, 964 char __b19, char __b18, char __b17, char __b16, 965 char __b15, char __b14, char __b13, char __b12, 966 char __b11, char __b10, char __b09, char __b08, 967 char __b07, char __b06, char __b05, char __b04, 968 char __b03, char __b02, char __b01, char __b00) 969 { 970 return (__m256i)(__v32qi){ 971 __b00, __b01, __b02, __b03, __b04, __b05, __b06, __b07, 972 __b08, __b09, __b10, __b11, __b12, __b13, __b14, __b15, 973 __b16, __b17, __b18, __b19, __b20, __b21, __b22, __b23, 974 __b24, __b25, __b26, __b27, __b28, __b29, __b30, __b31 975 }; 976 } 977 978 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 979 _mm256_set_epi64x(long long __a, long long __b, long long __c, long long __d) 980 { 981 return (__m256i)(__v4di){ __d, __c, __b, __a }; 982 } 983 984 /* Create vectors with elements in reverse order */ 985 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 986 _mm256_setr_pd(double __a, double __b, double __c, double __d) 987 { 988 return (__m256d){ __a, __b, __c, __d }; 989 } 990 991 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 992 _mm256_setr_ps(float __a, float __b, float __c, float __d, 993 float __e, float __f, float __g, float __h) 994 { 995 return (__m256){ __a, __b, __c, __d, __e, __f, __g, __h }; 996 } 997 998 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 999 _mm256_setr_epi32(int __i0, int __i1, int __i2, int __i3, 1000 int __i4, int __i5, int __i6, int __i7) 1001 { 1002 return (__m256i)(__v8si){ __i0, __i1, __i2, __i3, __i4, __i5, __i6, __i7 }; 1003 } 1004 1005 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1006 _mm256_setr_epi16(short __w15, short __w14, short __w13, short __w12, 1007 short __w11, short __w10, short __w09, short __w08, 1008 short __w07, short __w06, short __w05, short __w04, 1009 short __w03, short __w02, short __w01, short __w00) 1010 { 1011 return (__m256i)(__v16hi){ __w15, __w14, __w13, __w12, __w11, __w10, __w09, 1012 __w08, __w07, __w06, __w05, __w04, __w03, __w02, __w01, __w00 }; 1013 } 1014 1015 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1016 _mm256_setr_epi8(char __b31, char __b30, char __b29, char __b28, 1017 char __b27, char __b26, char __b25, char __b24, 1018 char __b23, char __b22, char __b21, char __b20, 1019 char __b19, char __b18, char __b17, char __b16, 1020 char __b15, char __b14, char __b13, char __b12, 1021 char __b11, char __b10, char __b09, char __b08, 1022 char __b07, char __b06, char __b05, char __b04, 1023 char __b03, char __b02, char __b01, char __b00) 1024 { 1025 return (__m256i)(__v32qi){ 1026 __b31, __b30, __b29, __b28, __b27, __b26, __b25, __b24, 1027 __b23, __b22, __b21, __b20, __b19, __b18, __b17, __b16, 1028 __b15, __b14, __b13, __b12, __b11, __b10, __b09, __b08, 1029 __b07, __b06, __b05, __b04, __b03, __b02, __b01, __b00 }; 1030 } 1031 1032 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1033 _mm256_setr_epi64x(long long __a, long long __b, long long __c, long long __d) 1034 { 1035 return (__m256i)(__v4di){ __a, __b, __c, __d }; 1036 } 1037 1038 /* Create vectors with repeated elements */ 1039 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1040 _mm256_set1_pd(double __w) 1041 { 1042 return (__m256d){ __w, __w, __w, __w }; 1043 } 1044 1045 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1046 _mm256_set1_ps(float __w) 1047 { 1048 return (__m256){ __w, __w, __w, __w, __w, __w, __w, __w }; 1049 } 1050 1051 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1052 _mm256_set1_epi32(int __i) 1053 { 1054 return (__m256i)(__v8si){ __i, __i, __i, __i, __i, __i, __i, __i }; 1055 } 1056 1057 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1058 _mm256_set1_epi16(short __w) 1059 { 1060 return (__m256i)(__v16hi){ __w, __w, __w, __w, __w, __w, __w, __w, __w, __w, 1061 __w, __w, __w, __w, __w, __w }; 1062 } 1063 1064 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1065 _mm256_set1_epi8(char __b) 1066 { 1067 return (__m256i)(__v32qi){ __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, 1068 __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, 1069 __b, __b, __b, __b, __b, __b, __b }; 1070 } 1071 1072 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1073 _mm256_set1_epi64x(long long __q) 1074 { 1075 return (__m256i)(__v4di){ __q, __q, __q, __q }; 1076 } 1077 1078 /* Create __zeroed vectors */ 1079 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1080 _mm256_setzero_pd(void) 1081 { 1082 return (__m256d){ 0, 0, 0, 0 }; 1083 } 1084 1085 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1086 _mm256_setzero_ps(void) 1087 { 1088 return (__m256){ 0, 0, 0, 0, 0, 0, 0, 0 }; 1089 } 1090 1091 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1092 _mm256_setzero_si256(void) 1093 { 1094 return (__m256i){ 0LL, 0LL, 0LL, 0LL }; 1095 } 1096 1097 /* Cast between vector types */ 1098 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1099 _mm256_castpd_ps(__m256d __a) 1100 { 1101 return (__m256)__a; 1102 } 1103 1104 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1105 _mm256_castpd_si256(__m256d __a) 1106 { 1107 return (__m256i)__a; 1108 } 1109 1110 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1111 _mm256_castps_pd(__m256 __a) 1112 { 1113 return (__m256d)__a; 1114 } 1115 1116 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1117 _mm256_castps_si256(__m256 __a) 1118 { 1119 return (__m256i)__a; 1120 } 1121 1122 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1123 _mm256_castsi256_ps(__m256i __a) 1124 { 1125 return (__m256)__a; 1126 } 1127 1128 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1129 _mm256_castsi256_pd(__m256i __a) 1130 { 1131 return (__m256d)__a; 1132 } 1133 1134 static __inline __m128d __attribute__((__always_inline__, __nodebug__)) 1135 _mm256_castpd256_pd128(__m256d __a) 1136 { 1137 return __builtin_shufflevector(__a, __a, 0, 1); 1138 } 1139 1140 static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 1141 _mm256_castps256_ps128(__m256 __a) 1142 { 1143 return __builtin_shufflevector(__a, __a, 0, 1, 2, 3); 1144 } 1145 1146 static __inline __m128i __attribute__((__always_inline__, __nodebug__)) 1147 _mm256_castsi256_si128(__m256i __a) 1148 { 1149 return __builtin_shufflevector(__a, __a, 0, 1); 1150 } 1151 1152 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1153 _mm256_castpd128_pd256(__m128d __a) 1154 { 1155 return __builtin_shufflevector(__a, __a, 0, 1, -1, -1); 1156 } 1157 1158 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1159 _mm256_castps128_ps256(__m128 __a) 1160 { 1161 return __builtin_shufflevector(__a, __a, 0, 1, 2, 3, -1, -1, -1, -1); 1162 } 1163 1164 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1165 _mm256_castsi128_si256(__m128i __a) 1166 { 1167 return __builtin_shufflevector(__a, __a, 0, 1, -1, -1); 1168 } 1169 1170 /* SIMD load ops (unaligned) */ 1171 static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1172 _mm256_loadu2_m128(float const *__addr_hi, float const *__addr_lo) 1173 { 1174 struct __loadu_ps { 1175 __m128 __v; 1176 } __attribute__((__packed__, __may_alias__)); 1177 1178 __m256 __v256 = _mm256_castps128_ps256(((struct __loadu_ps*)__addr_lo)->__v); 1179 return _mm256_insertf128_ps(__v256, ((struct __loadu_ps*)__addr_hi)->__v, 1); 1180 } 1181 1182 static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1183 _mm256_loadu2_m128d(double const *__addr_hi, double const *__addr_lo) 1184 { 1185 struct __loadu_pd { 1186 __m128d __v; 1187 } __attribute__((__packed__, __may_alias__)); 1188 1189 __m256d __v256 = _mm256_castpd128_pd256(((struct __loadu_pd*)__addr_lo)->__v); 1190 return _mm256_insertf128_pd(__v256, ((struct __loadu_pd*)__addr_hi)->__v, 1); 1191 } 1192 1193 static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1194 _mm256_loadu2_m128i(__m128i const *__addr_hi, __m128i const *__addr_lo) 1195 { 1196 struct __loadu_si128 { 1197 __m128i __v; 1198 } __attribute__((packed, may_alias)); 1199 __m256i __v256 = _mm256_castsi128_si256( 1200 ((struct __loadu_si128*)__addr_lo)->__v); 1201 return _mm256_insertf128_si256(__v256, 1202 ((struct __loadu_si128*)__addr_hi)->__v, 1); 1203 } 1204 1205 /* SIMD store ops (unaligned) */ 1206 static __inline void __attribute__((__always_inline__, __nodebug__)) 1207 _mm256_storeu2_m128(float *__addr_hi, float *__addr_lo, __m256 __a) 1208 { 1209 __m128 __v128; 1210 1211 __v128 = _mm256_castps256_ps128(__a); 1212 __builtin_ia32_storeups(__addr_lo, __v128); 1213 __v128 = _mm256_extractf128_ps(__a, 1); 1214 __builtin_ia32_storeups(__addr_hi, __v128); 1215 } 1216 1217 static __inline void __attribute__((__always_inline__, __nodebug__)) 1218 _mm256_storeu2_m128d(double *__addr_hi, double *__addr_lo, __m256d __a) 1219 { 1220 __m128d __v128; 1221 1222 __v128 = _mm256_castpd256_pd128(__a); 1223 __builtin_ia32_storeupd(__addr_lo, __v128); 1224 __v128 = _mm256_extractf128_pd(__a, 1); 1225 __builtin_ia32_storeupd(__addr_hi, __v128); 1226 } 1227 1228 static __inline void __attribute__((__always_inline__, __nodebug__)) 1229 _mm256_storeu2_m128i(__m128i *__addr_hi, __m128i *__addr_lo, __m256i __a) 1230 { 1231 __m128i __v128; 1232 1233 __v128 = _mm256_castsi256_si128(__a); 1234 __builtin_ia32_storedqu((char *)__addr_lo, (__v16qi)__v128); 1235 __v128 = _mm256_extractf128_si256(__a, 1); 1236 __builtin_ia32_storedqu((char *)__addr_hi, (__v16qi)__v128); 1237 } 1238 1239 #endif /* __AVXINTRIN_H */ 1240