avxintrin.h revision 249423
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 28typedef double __v4df __attribute__ ((__vector_size__ (32))); 29typedef float __v8sf __attribute__ ((__vector_size__ (32))); 30typedef long long __v4di __attribute__ ((__vector_size__ (32))); 31typedef int __v8si __attribute__ ((__vector_size__ (32))); 32typedef short __v16hi __attribute__ ((__vector_size__ (32))); 33typedef char __v32qi __attribute__ ((__vector_size__ (32))); 34 35typedef float __m256 __attribute__ ((__vector_size__ (32))); 36typedef double __m256d __attribute__((__vector_size__(32))); 37typedef long long __m256i __attribute__((__vector_size__(32))); 38 39/* Arithmetic */ 40static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 41_mm256_add_pd(__m256d __a, __m256d __b) 42{ 43 return __a+__b; 44} 45 46static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 47_mm256_add_ps(__m256 __a, __m256 __b) 48{ 49 return __a+__b; 50} 51 52static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 53_mm256_sub_pd(__m256d __a, __m256d __b) 54{ 55 return __a-__b; 56} 57 58static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 59_mm256_sub_ps(__m256 __a, __m256 __b) 60{ 61 return __a-__b; 62} 63 64static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 65_mm256_addsub_pd(__m256d __a, __m256d __b) 66{ 67 return (__m256d)__builtin_ia32_addsubpd256((__v4df)__a, (__v4df)__b); 68} 69 70static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 71_mm256_addsub_ps(__m256 __a, __m256 __b) 72{ 73 return (__m256)__builtin_ia32_addsubps256((__v8sf)__a, (__v8sf)__b); 74} 75 76static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 77_mm256_div_pd(__m256d __a, __m256d __b) 78{ 79 return __a / __b; 80} 81 82static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 83_mm256_div_ps(__m256 __a, __m256 __b) 84{ 85 return __a / __b; 86} 87 88static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 89_mm256_max_pd(__m256d __a, __m256d __b) 90{ 91 return (__m256d)__builtin_ia32_maxpd256((__v4df)__a, (__v4df)__b); 92} 93 94static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 95_mm256_max_ps(__m256 __a, __m256 __b) 96{ 97 return (__m256)__builtin_ia32_maxps256((__v8sf)__a, (__v8sf)__b); 98} 99 100static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 101_mm256_min_pd(__m256d __a, __m256d __b) 102{ 103 return (__m256d)__builtin_ia32_minpd256((__v4df)__a, (__v4df)__b); 104} 105 106static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 107_mm256_min_ps(__m256 __a, __m256 __b) 108{ 109 return (__m256)__builtin_ia32_minps256((__v8sf)__a, (__v8sf)__b); 110} 111 112static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 113_mm256_mul_pd(__m256d __a, __m256d __b) 114{ 115 return __a * __b; 116} 117 118static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 119_mm256_mul_ps(__m256 __a, __m256 __b) 120{ 121 return __a * __b; 122} 123 124static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 125_mm256_sqrt_pd(__m256d __a) 126{ 127 return (__m256d)__builtin_ia32_sqrtpd256((__v4df)__a); 128} 129 130static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 131_mm256_sqrt_ps(__m256 __a) 132{ 133 return (__m256)__builtin_ia32_sqrtps256((__v8sf)__a); 134} 135 136static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 137_mm256_rsqrt_ps(__m256 __a) 138{ 139 return (__m256)__builtin_ia32_rsqrtps256((__v8sf)__a); 140} 141 142static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 143_mm256_rcp_ps(__m256 __a) 144{ 145 return (__m256)__builtin_ia32_rcpps256((__v8sf)__a); 146} 147 148#define _mm256_round_pd(V, M) __extension__ ({ \ 149 __m256d __V = (V); \ 150 (__m256d)__builtin_ia32_roundpd256((__v4df)__V, (M)); }) 151 152#define _mm256_round_ps(V, M) __extension__ ({ \ 153 __m256 __V = (V); \ 154 (__m256)__builtin_ia32_roundps256((__v8sf)__V, (M)); }) 155 156#define _mm256_ceil_pd(V) _mm256_round_pd((V), _MM_FROUND_CEIL) 157#define _mm256_floor_pd(V) _mm256_round_pd((V), _MM_FROUND_FLOOR) 158#define _mm256_ceil_ps(V) _mm256_round_ps((V), _MM_FROUND_CEIL) 159#define _mm256_floor_ps(V) _mm256_round_ps((V), _MM_FROUND_FLOOR) 160 161/* Logical */ 162static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 163_mm256_and_pd(__m256d __a, __m256d __b) 164{ 165 return (__m256d)((__v4di)__a & (__v4di)__b); 166} 167 168static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 169_mm256_and_ps(__m256 __a, __m256 __b) 170{ 171 return (__m256)((__v8si)__a & (__v8si)__b); 172} 173 174static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 175_mm256_andnot_pd(__m256d __a, __m256d __b) 176{ 177 return (__m256d)(~(__v4di)__a & (__v4di)__b); 178} 179 180static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 181_mm256_andnot_ps(__m256 __a, __m256 __b) 182{ 183 return (__m256)(~(__v8si)__a & (__v8si)__b); 184} 185 186static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 187_mm256_or_pd(__m256d __a, __m256d __b) 188{ 189 return (__m256d)((__v4di)__a | (__v4di)__b); 190} 191 192static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 193_mm256_or_ps(__m256 __a, __m256 __b) 194{ 195 return (__m256)((__v8si)__a | (__v8si)__b); 196} 197 198static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 199_mm256_xor_pd(__m256d __a, __m256d __b) 200{ 201 return (__m256d)((__v4di)__a ^ (__v4di)__b); 202} 203 204static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 205_mm256_xor_ps(__m256 __a, __m256 __b) 206{ 207 return (__m256)((__v8si)__a ^ (__v8si)__b); 208} 209 210/* Horizontal arithmetic */ 211static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 212_mm256_hadd_pd(__m256d __a, __m256d __b) 213{ 214 return (__m256d)__builtin_ia32_haddpd256((__v4df)__a, (__v4df)__b); 215} 216 217static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 218_mm256_hadd_ps(__m256 __a, __m256 __b) 219{ 220 return (__m256)__builtin_ia32_haddps256((__v8sf)__a, (__v8sf)__b); 221} 222 223static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 224_mm256_hsub_pd(__m256d __a, __m256d __b) 225{ 226 return (__m256d)__builtin_ia32_hsubpd256((__v4df)__a, (__v4df)__b); 227} 228 229static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 230_mm256_hsub_ps(__m256 __a, __m256 __b) 231{ 232 return (__m256)__builtin_ia32_hsubps256((__v8sf)__a, (__v8sf)__b); 233} 234 235/* Vector permutations */ 236static __inline __m128d __attribute__((__always_inline__, __nodebug__)) 237_mm_permutevar_pd(__m128d __a, __m128i __c) 238{ 239 return (__m128d)__builtin_ia32_vpermilvarpd((__v2df)__a, (__v2di)__c); 240} 241 242static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 243_mm256_permutevar_pd(__m256d __a, __m256i __c) 244{ 245 return (__m256d)__builtin_ia32_vpermilvarpd256((__v4df)__a, (__v4di)__c); 246} 247 248static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 249_mm_permutevar_ps(__m128 __a, __m128i __c) 250{ 251 return (__m128)__builtin_ia32_vpermilvarps((__v4sf)__a, (__v4si)__c); 252} 253 254static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 255_mm256_permutevar_ps(__m256 __a, __m256i __c) 256{ 257 return (__m256)__builtin_ia32_vpermilvarps256((__v8sf)__a, 258 (__v8si)__c); 259} 260 261#define _mm_permute_pd(A, C) __extension__ ({ \ 262 __m128d __A = (A); \ 263 (__m128d)__builtin_shufflevector((__v2df)__A, (__v2df) _mm_setzero_pd(), \ 264 (C) & 0x1, ((C) & 0x2) >> 1); }) 265 266#define _mm256_permute_pd(A, C) __extension__ ({ \ 267 __m256d __A = (A); \ 268 (__m256d)__builtin_shufflevector((__v4df)__A, (__v4df) _mm256_setzero_pd(), \ 269 (C) & 0x1, ((C) & 0x2) >> 1, \ 270 2 + (((C) & 0x4) >> 2), \ 271 2 + (((C) & 0x8) >> 3)); }) 272 273#define _mm_permute_ps(A, C) __extension__ ({ \ 274 __m128 __A = (A); \ 275 (__m128)__builtin_shufflevector((__v4sf)__A, (__v4sf) _mm_setzero_ps(), \ 276 (C) & 0x3, ((C) & 0xc) >> 2, \ 277 ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6); }) 278 279#define _mm256_permute_ps(A, C) __extension__ ({ \ 280 __m256 __A = (A); \ 281 (__m256)__builtin_shufflevector((__v8sf)__A, (__v8sf) _mm256_setzero_ps(), \ 282 (C) & 0x3, ((C) & 0xc) >> 2, \ 283 ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6, \ 284 4 + (((C) & 0x03) >> 0), \ 285 4 + (((C) & 0x0c) >> 2), \ 286 4 + (((C) & 0x30) >> 4), \ 287 4 + (((C) & 0xc0) >> 6)); }) 288 289#define _mm256_permute2f128_pd(V1, V2, M) __extension__ ({ \ 290 __m256d __V1 = (V1); \ 291 __m256d __V2 = (V2); \ 292 (__m256d)__builtin_ia32_vperm2f128_pd256((__v4df)__V1, (__v4df)__V2, (M)); }) 293 294#define _mm256_permute2f128_ps(V1, V2, M) __extension__ ({ \ 295 __m256 __V1 = (V1); \ 296 __m256 __V2 = (V2); \ 297 (__m256)__builtin_ia32_vperm2f128_ps256((__v8sf)__V1, (__v8sf)__V2, (M)); }) 298 299#define _mm256_permute2f128_si256(V1, V2, M) __extension__ ({ \ 300 __m256i __V1 = (V1); \ 301 __m256i __V2 = (V2); \ 302 (__m256i)__builtin_ia32_vperm2f128_si256((__v8si)__V1, (__v8si)__V2, (M)); }) 303 304/* Vector Blend */ 305#define _mm256_blend_pd(V1, V2, M) __extension__ ({ \ 306 __m256d __V1 = (V1); \ 307 __m256d __V2 = (V2); \ 308 (__m256d)__builtin_ia32_blendpd256((__v4df)__V1, (__v4df)__V2, (M)); }) 309 310#define _mm256_blend_ps(V1, V2, M) __extension__ ({ \ 311 __m256 __V1 = (V1); \ 312 __m256 __V2 = (V2); \ 313 (__m256)__builtin_ia32_blendps256((__v8sf)__V1, (__v8sf)__V2, (M)); }) 314 315static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 316_mm256_blendv_pd(__m256d __a, __m256d __b, __m256d __c) 317{ 318 return (__m256d)__builtin_ia32_blendvpd256( 319 (__v4df)__a, (__v4df)__b, (__v4df)__c); 320} 321 322static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 323_mm256_blendv_ps(__m256 __a, __m256 __b, __m256 __c) 324{ 325 return (__m256)__builtin_ia32_blendvps256( 326 (__v8sf)__a, (__v8sf)__b, (__v8sf)__c); 327} 328 329/* Vector Dot Product */ 330#define _mm256_dp_ps(V1, V2, M) __extension__ ({ \ 331 __m256 __V1 = (V1); \ 332 __m256 __V2 = (V2); \ 333 (__m256)__builtin_ia32_dpps256((__v8sf)__V1, (__v8sf)__V2, (M)); }) 334 335/* Vector shuffle */ 336#define _mm256_shuffle_ps(a, b, mask) __extension__ ({ \ 337 __m256 __a = (a); \ 338 __m256 __b = (b); \ 339 (__m256)__builtin_shufflevector((__v8sf)__a, (__v8sf)__b, \ 340 (mask) & 0x3, ((mask) & 0xc) >> 2, \ 341 (((mask) & 0x30) >> 4) + 8, (((mask) & 0xc0) >> 6) + 8, \ 342 ((mask) & 0x3) + 4, (((mask) & 0xc) >> 2) + 4, \ 343 (((mask) & 0x30) >> 4) + 12, (((mask) & 0xc0) >> 6) + 12); }) 344 345#define _mm256_shuffle_pd(a, b, mask) __extension__ ({ \ 346 __m256d __a = (a); \ 347 __m256d __b = (b); \ 348 (__m256d)__builtin_shufflevector((__v4df)__a, (__v4df)__b, \ 349 (mask) & 0x1, \ 350 (((mask) & 0x2) >> 1) + 4, \ 351 (((mask) & 0x4) >> 2) + 2, \ 352 (((mask) & 0x8) >> 3) + 6); }) 353 354/* Compare */ 355#define _CMP_EQ_OQ 0x00 /* Equal (ordered, non-signaling) */ 356#define _CMP_LT_OS 0x01 /* Less-than (ordered, signaling) */ 357#define _CMP_LE_OS 0x02 /* Less-than-or-equal (ordered, signaling) */ 358#define _CMP_UNORD_Q 0x03 /* Unordered (non-signaling) */ 359#define _CMP_NEQ_UQ 0x04 /* Not-equal (unordered, non-signaling) */ 360#define _CMP_NLT_US 0x05 /* Not-less-than (unordered, signaling) */ 361#define _CMP_NLE_US 0x06 /* Not-less-than-or-equal (unordered, signaling) */ 362#define _CMP_ORD_Q 0x07 /* Ordered (nonsignaling) */ 363#define _CMP_EQ_UQ 0x08 /* Equal (unordered, non-signaling) */ 364#define _CMP_NGE_US 0x09 /* Not-greater-than-or-equal (unord, signaling) */ 365#define _CMP_NGT_US 0x0a /* Not-greater-than (unordered, signaling) */ 366#define _CMP_FALSE_OQ 0x0b /* False (ordered, non-signaling) */ 367#define _CMP_NEQ_OQ 0x0c /* Not-equal (ordered, non-signaling) */ 368#define _CMP_GE_OS 0x0d /* Greater-than-or-equal (ordered, signaling) */ 369#define _CMP_GT_OS 0x0e /* Greater-than (ordered, signaling) */ 370#define _CMP_TRUE_UQ 0x0f /* True (unordered, non-signaling) */ 371#define _CMP_EQ_OS 0x10 /* Equal (ordered, signaling) */ 372#define _CMP_LT_OQ 0x11 /* Less-than (ordered, non-signaling) */ 373#define _CMP_LE_OQ 0x12 /* Less-than-or-equal (ordered, non-signaling) */ 374#define _CMP_UNORD_S 0x13 /* Unordered (signaling) */ 375#define _CMP_NEQ_US 0x14 /* Not-equal (unordered, signaling) */ 376#define _CMP_NLT_UQ 0x15 /* Not-less-than (unordered, non-signaling) */ 377#define _CMP_NLE_UQ 0x16 /* Not-less-than-or-equal (unord, non-signaling) */ 378#define _CMP_ORD_S 0x17 /* Ordered (signaling) */ 379#define _CMP_EQ_US 0x18 /* Equal (unordered, signaling) */ 380#define _CMP_NGE_UQ 0x19 /* Not-greater-than-or-equal (unord, non-sign) */ 381#define _CMP_NGT_UQ 0x1a /* Not-greater-than (unordered, non-signaling) */ 382#define _CMP_FALSE_OS 0x1b /* False (ordered, signaling) */ 383#define _CMP_NEQ_OS 0x1c /* Not-equal (ordered, signaling) */ 384#define _CMP_GE_OQ 0x1d /* Greater-than-or-equal (ordered, non-signaling) */ 385#define _CMP_GT_OQ 0x1e /* Greater-than (ordered, non-signaling) */ 386#define _CMP_TRUE_US 0x1f /* True (unordered, signaling) */ 387 388#define _mm_cmp_pd(a, b, c) __extension__ ({ \ 389 __m128d __a = (a); \ 390 __m128d __b = (b); \ 391 (__m128d)__builtin_ia32_cmppd((__v2df)__a, (__v2df)__b, (c)); }) 392 393#define _mm_cmp_ps(a, b, c) __extension__ ({ \ 394 __m128 __a = (a); \ 395 __m128 __b = (b); \ 396 (__m128)__builtin_ia32_cmpps((__v4sf)__a, (__v4sf)__b, (c)); }) 397 398#define _mm256_cmp_pd(a, b, c) __extension__ ({ \ 399 __m256d __a = (a); \ 400 __m256d __b = (b); \ 401 (__m256d)__builtin_ia32_cmppd256((__v4df)__a, (__v4df)__b, (c)); }) 402 403#define _mm256_cmp_ps(a, b, c) __extension__ ({ \ 404 __m256 __a = (a); \ 405 __m256 __b = (b); \ 406 (__m256)__builtin_ia32_cmpps256((__v8sf)__a, (__v8sf)__b, (c)); }) 407 408#define _mm_cmp_sd(a, b, c) __extension__ ({ \ 409 __m128d __a = (a); \ 410 __m128d __b = (b); \ 411 (__m128d)__builtin_ia32_cmpsd((__v2df)__a, (__v2df)__b, (c)); }) 412 413#define _mm_cmp_ss(a, b, c) __extension__ ({ \ 414 __m128 __a = (a); \ 415 __m128 __b = (b); \ 416 (__m128)__builtin_ia32_cmpss((__v4sf)__a, (__v4sf)__b, (c)); }) 417 418/* Vector extract */ 419#define _mm256_extractf128_pd(A, O) __extension__ ({ \ 420 __m256d __A = (A); \ 421 (__m128d)__builtin_ia32_vextractf128_pd256((__v4df)__A, (O)); }) 422 423#define _mm256_extractf128_ps(A, O) __extension__ ({ \ 424 __m256 __A = (A); \ 425 (__m128)__builtin_ia32_vextractf128_ps256((__v8sf)__A, (O)); }) 426 427#define _mm256_extractf128_si256(A, O) __extension__ ({ \ 428 __m256i __A = (A); \ 429 (__m128i)__builtin_ia32_vextractf128_si256((__v8si)__A, (O)); }) 430 431static __inline int __attribute__((__always_inline__, __nodebug__)) 432_mm256_extract_epi32(__m256i __a, int const __imm) 433{ 434 __v8si __b = (__v8si)__a; 435 return __b[__imm]; 436} 437 438static __inline int __attribute__((__always_inline__, __nodebug__)) 439_mm256_extract_epi16(__m256i __a, int const __imm) 440{ 441 __v16hi __b = (__v16hi)__a; 442 return __b[__imm]; 443} 444 445static __inline int __attribute__((__always_inline__, __nodebug__)) 446_mm256_extract_epi8(__m256i __a, int const __imm) 447{ 448 __v32qi __b = (__v32qi)__a; 449 return __b[__imm]; 450} 451 452#ifdef __x86_64__ 453static __inline long long __attribute__((__always_inline__, __nodebug__)) 454_mm256_extract_epi64(__m256i __a, const int __imm) 455{ 456 __v4di __b = (__v4di)__a; 457 return __b[__imm]; 458} 459#endif 460 461/* Vector insert */ 462#define _mm256_insertf128_pd(V1, V2, O) __extension__ ({ \ 463 __m256d __V1 = (V1); \ 464 __m128d __V2 = (V2); \ 465 (__m256d)__builtin_ia32_vinsertf128_pd256((__v4df)__V1, (__v2df)__V2, (O)); }) 466 467#define _mm256_insertf128_ps(V1, V2, O) __extension__ ({ \ 468 __m256 __V1 = (V1); \ 469 __m128 __V2 = (V2); \ 470 (__m256)__builtin_ia32_vinsertf128_ps256((__v8sf)__V1, (__v4sf)__V2, (O)); }) 471 472#define _mm256_insertf128_si256(V1, V2, O) __extension__ ({ \ 473 __m256i __V1 = (V1); \ 474 __m128i __V2 = (V2); \ 475 (__m256i)__builtin_ia32_vinsertf128_si256((__v8si)__V1, (__v4si)__V2, (O)); }) 476 477static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 478_mm256_insert_epi32(__m256i __a, int __b, int const __imm) 479{ 480 __v8si __c = (__v8si)__a; 481 __c[__imm & 7] = __b; 482 return (__m256i)__c; 483} 484 485static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 486_mm256_insert_epi16(__m256i __a, int __b, int const __imm) 487{ 488 __v16hi __c = (__v16hi)__a; 489 __c[__imm & 15] = __b; 490 return (__m256i)__c; 491} 492 493static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 494_mm256_insert_epi8(__m256i __a, int __b, int const __imm) 495{ 496 __v32qi __c = (__v32qi)__a; 497 __c[__imm & 31] = __b; 498 return (__m256i)__c; 499} 500 501#ifdef __x86_64__ 502static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 503_mm256_insert_epi64(__m256i __a, int __b, int const __imm) 504{ 505 __v4di __c = (__v4di)__a; 506 __c[__imm & 3] = __b; 507 return (__m256i)__c; 508} 509#endif 510 511/* Conversion */ 512static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 513_mm256_cvtepi32_pd(__m128i __a) 514{ 515 return (__m256d)__builtin_ia32_cvtdq2pd256((__v4si) __a); 516} 517 518static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 519_mm256_cvtepi32_ps(__m256i __a) 520{ 521 return (__m256)__builtin_ia32_cvtdq2ps256((__v8si) __a); 522} 523 524static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 525_mm256_cvtpd_ps(__m256d __a) 526{ 527 return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) __a); 528} 529 530static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 531_mm256_cvtps_epi32(__m256 __a) 532{ 533 return (__m256i)__builtin_ia32_cvtps2dq256((__v8sf) __a); 534} 535 536static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 537_mm256_cvtps_pd(__m128 __a) 538{ 539 return (__m256d)__builtin_ia32_cvtps2pd256((__v4sf) __a); 540} 541 542static __inline __m128i __attribute__((__always_inline__, __nodebug__)) 543_mm256_cvttpd_epi32(__m256d __a) 544{ 545 return (__m128i)__builtin_ia32_cvttpd2dq256((__v4df) __a); 546} 547 548static __inline __m128i __attribute__((__always_inline__, __nodebug__)) 549_mm256_cvtpd_epi32(__m256d __a) 550{ 551 return (__m128i)__builtin_ia32_cvtpd2dq256((__v4df) __a); 552} 553 554static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 555_mm256_cvttps_epi32(__m256 __a) 556{ 557 return (__m256i)__builtin_ia32_cvttps2dq256((__v8sf) __a); 558} 559 560/* Vector replicate */ 561static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 562_mm256_movehdup_ps(__m256 __a) 563{ 564 return __builtin_shufflevector(__a, __a, 1, 1, 3, 3, 5, 5, 7, 7); 565} 566 567static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 568_mm256_moveldup_ps(__m256 __a) 569{ 570 return __builtin_shufflevector(__a, __a, 0, 0, 2, 2, 4, 4, 6, 6); 571} 572 573static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 574_mm256_movedup_pd(__m256d __a) 575{ 576 return __builtin_shufflevector(__a, __a, 0, 0, 2, 2); 577} 578 579/* Unpack and Interleave */ 580static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 581_mm256_unpackhi_pd(__m256d __a, __m256d __b) 582{ 583 return __builtin_shufflevector(__a, __b, 1, 5, 1+2, 5+2); 584} 585 586static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 587_mm256_unpacklo_pd(__m256d __a, __m256d __b) 588{ 589 return __builtin_shufflevector(__a, __b, 0, 4, 0+2, 4+2); 590} 591 592static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 593_mm256_unpackhi_ps(__m256 __a, __m256 __b) 594{ 595 return __builtin_shufflevector(__a, __b, 2, 10, 2+1, 10+1, 6, 14, 6+1, 14+1); 596} 597 598static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 599_mm256_unpacklo_ps(__m256 __a, __m256 __b) 600{ 601 return __builtin_shufflevector(__a, __b, 0, 8, 0+1, 8+1, 4, 12, 4+1, 12+1); 602} 603 604/* Bit Test */ 605static __inline int __attribute__((__always_inline__, __nodebug__)) 606_mm_testz_pd(__m128d __a, __m128d __b) 607{ 608 return __builtin_ia32_vtestzpd((__v2df)__a, (__v2df)__b); 609} 610 611static __inline int __attribute__((__always_inline__, __nodebug__)) 612_mm_testc_pd(__m128d __a, __m128d __b) 613{ 614 return __builtin_ia32_vtestcpd((__v2df)__a, (__v2df)__b); 615} 616 617static __inline int __attribute__((__always_inline__, __nodebug__)) 618_mm_testnzc_pd(__m128d __a, __m128d __b) 619{ 620 return __builtin_ia32_vtestnzcpd((__v2df)__a, (__v2df)__b); 621} 622 623static __inline int __attribute__((__always_inline__, __nodebug__)) 624_mm_testz_ps(__m128 __a, __m128 __b) 625{ 626 return __builtin_ia32_vtestzps((__v4sf)__a, (__v4sf)__b); 627} 628 629static __inline int __attribute__((__always_inline__, __nodebug__)) 630_mm_testc_ps(__m128 __a, __m128 __b) 631{ 632 return __builtin_ia32_vtestcps((__v4sf)__a, (__v4sf)__b); 633} 634 635static __inline int __attribute__((__always_inline__, __nodebug__)) 636_mm_testnzc_ps(__m128 __a, __m128 __b) 637{ 638 return __builtin_ia32_vtestnzcps((__v4sf)__a, (__v4sf)__b); 639} 640 641static __inline int __attribute__((__always_inline__, __nodebug__)) 642_mm256_testz_pd(__m256d __a, __m256d __b) 643{ 644 return __builtin_ia32_vtestzpd256((__v4df)__a, (__v4df)__b); 645} 646 647static __inline int __attribute__((__always_inline__, __nodebug__)) 648_mm256_testc_pd(__m256d __a, __m256d __b) 649{ 650 return __builtin_ia32_vtestcpd256((__v4df)__a, (__v4df)__b); 651} 652 653static __inline int __attribute__((__always_inline__, __nodebug__)) 654_mm256_testnzc_pd(__m256d __a, __m256d __b) 655{ 656 return __builtin_ia32_vtestnzcpd256((__v4df)__a, (__v4df)__b); 657} 658 659static __inline int __attribute__((__always_inline__, __nodebug__)) 660_mm256_testz_ps(__m256 __a, __m256 __b) 661{ 662 return __builtin_ia32_vtestzps256((__v8sf)__a, (__v8sf)__b); 663} 664 665static __inline int __attribute__((__always_inline__, __nodebug__)) 666_mm256_testc_ps(__m256 __a, __m256 __b) 667{ 668 return __builtin_ia32_vtestcps256((__v8sf)__a, (__v8sf)__b); 669} 670 671static __inline int __attribute__((__always_inline__, __nodebug__)) 672_mm256_testnzc_ps(__m256 __a, __m256 __b) 673{ 674 return __builtin_ia32_vtestnzcps256((__v8sf)__a, (__v8sf)__b); 675} 676 677static __inline int __attribute__((__always_inline__, __nodebug__)) 678_mm256_testz_si256(__m256i __a, __m256i __b) 679{ 680 return __builtin_ia32_ptestz256((__v4di)__a, (__v4di)__b); 681} 682 683static __inline int __attribute__((__always_inline__, __nodebug__)) 684_mm256_testc_si256(__m256i __a, __m256i __b) 685{ 686 return __builtin_ia32_ptestc256((__v4di)__a, (__v4di)__b); 687} 688 689static __inline int __attribute__((__always_inline__, __nodebug__)) 690_mm256_testnzc_si256(__m256i __a, __m256i __b) 691{ 692 return __builtin_ia32_ptestnzc256((__v4di)__a, (__v4di)__b); 693} 694 695/* Vector extract sign mask */ 696static __inline int __attribute__((__always_inline__, __nodebug__)) 697_mm256_movemask_pd(__m256d __a) 698{ 699 return __builtin_ia32_movmskpd256((__v4df)__a); 700} 701 702static __inline int __attribute__((__always_inline__, __nodebug__)) 703_mm256_movemask_ps(__m256 __a) 704{ 705 return __builtin_ia32_movmskps256((__v8sf)__a); 706} 707 708/* Vector __zero */ 709static __inline void __attribute__((__always_inline__, __nodebug__)) 710_mm256_zeroall(void) 711{ 712 __builtin_ia32_vzeroall(); 713} 714 715static __inline void __attribute__((__always_inline__, __nodebug__)) 716_mm256_zeroupper(void) 717{ 718 __builtin_ia32_vzeroupper(); 719} 720 721/* Vector load with broadcast */ 722static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 723_mm_broadcast_ss(float const *__a) 724{ 725 return (__m128)__builtin_ia32_vbroadcastss(__a); 726} 727 728static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 729_mm256_broadcast_sd(double const *__a) 730{ 731 return (__m256d)__builtin_ia32_vbroadcastsd256(__a); 732} 733 734static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 735_mm256_broadcast_ss(float const *__a) 736{ 737 return (__m256)__builtin_ia32_vbroadcastss256(__a); 738} 739 740static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 741_mm256_broadcast_pd(__m128d const *__a) 742{ 743 return (__m256d)__builtin_ia32_vbroadcastf128_pd256(__a); 744} 745 746static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 747_mm256_broadcast_ps(__m128 const *__a) 748{ 749 return (__m256)__builtin_ia32_vbroadcastf128_ps256(__a); 750} 751 752/* SIMD load ops */ 753static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 754_mm256_load_pd(double const *__p) 755{ 756 return *(__m256d *)__p; 757} 758 759static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 760_mm256_load_ps(float const *__p) 761{ 762 return *(__m256 *)__p; 763} 764 765static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 766_mm256_loadu_pd(double const *__p) 767{ 768 struct __loadu_pd { 769 __m256d __v; 770 } __attribute__((packed, may_alias)); 771 return ((struct __loadu_pd*)__p)->__v; 772} 773 774static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 775_mm256_loadu_ps(float const *__p) 776{ 777 struct __loadu_ps { 778 __m256 __v; 779 } __attribute__((packed, may_alias)); 780 return ((struct __loadu_ps*)__p)->__v; 781} 782 783static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 784_mm256_load_si256(__m256i const *__p) 785{ 786 return *__p; 787} 788 789static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 790_mm256_loadu_si256(__m256i const *__p) 791{ 792 struct __loadu_si256 { 793 __m256i __v; 794 } __attribute__((packed, may_alias)); 795 return ((struct __loadu_si256*)__p)->__v; 796} 797 798static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 799_mm256_lddqu_si256(__m256i const *__p) 800{ 801 return (__m256i)__builtin_ia32_lddqu256((char const *)__p); 802} 803 804/* SIMD store ops */ 805static __inline void __attribute__((__always_inline__, __nodebug__)) 806_mm256_store_pd(double *__p, __m256d __a) 807{ 808 *(__m256d *)__p = __a; 809} 810 811static __inline void __attribute__((__always_inline__, __nodebug__)) 812_mm256_store_ps(float *__p, __m256 __a) 813{ 814 *(__m256 *)__p = __a; 815} 816 817static __inline void __attribute__((__always_inline__, __nodebug__)) 818_mm256_storeu_pd(double *__p, __m256d __a) 819{ 820 __builtin_ia32_storeupd256(__p, (__v4df)__a); 821} 822 823static __inline void __attribute__((__always_inline__, __nodebug__)) 824_mm256_storeu_ps(float *__p, __m256 __a) 825{ 826 __builtin_ia32_storeups256(__p, (__v8sf)__a); 827} 828 829static __inline void __attribute__((__always_inline__, __nodebug__)) 830_mm256_store_si256(__m256i *__p, __m256i __a) 831{ 832 *__p = __a; 833} 834 835static __inline void __attribute__((__always_inline__, __nodebug__)) 836_mm256_storeu_si256(__m256i *__p, __m256i __a) 837{ 838 __builtin_ia32_storedqu256((char *)__p, (__v32qi)__a); 839} 840 841/* Conditional load ops */ 842static __inline __m128d __attribute__((__always_inline__, __nodebug__)) 843_mm_maskload_pd(double const *__p, __m128d __m) 844{ 845 return (__m128d)__builtin_ia32_maskloadpd((const __v2df *)__p, (__v2df)__m); 846} 847 848static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 849_mm256_maskload_pd(double const *__p, __m256d __m) 850{ 851 return (__m256d)__builtin_ia32_maskloadpd256((const __v4df *)__p, 852 (__v4df)__m); 853} 854 855static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 856_mm_maskload_ps(float const *__p, __m128 __m) 857{ 858 return (__m128)__builtin_ia32_maskloadps((const __v4sf *)__p, (__v4sf)__m); 859} 860 861static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 862_mm256_maskload_ps(float const *__p, __m256 __m) 863{ 864 return (__m256)__builtin_ia32_maskloadps256((const __v8sf *)__p, (__v8sf)__m); 865} 866 867/* Conditional store ops */ 868static __inline void __attribute__((__always_inline__, __nodebug__)) 869_mm256_maskstore_ps(float *__p, __m256 __m, __m256 __a) 870{ 871 __builtin_ia32_maskstoreps256((__v8sf *)__p, (__v8sf)__m, (__v8sf)__a); 872} 873 874static __inline void __attribute__((__always_inline__, __nodebug__)) 875_mm_maskstore_pd(double *__p, __m128d __m, __m128d __a) 876{ 877 __builtin_ia32_maskstorepd((__v2df *)__p, (__v2df)__m, (__v2df)__a); 878} 879 880static __inline void __attribute__((__always_inline__, __nodebug__)) 881_mm256_maskstore_pd(double *__p, __m256d __m, __m256d __a) 882{ 883 __builtin_ia32_maskstorepd256((__v4df *)__p, (__v4df)__m, (__v4df)__a); 884} 885 886static __inline void __attribute__((__always_inline__, __nodebug__)) 887_mm_maskstore_ps(float *__p, __m128 __m, __m128 __a) 888{ 889 __builtin_ia32_maskstoreps((__v4sf *)__p, (__v4sf)__m, (__v4sf)__a); 890} 891 892/* Cacheability support ops */ 893static __inline void __attribute__((__always_inline__, __nodebug__)) 894_mm256_stream_si256(__m256i *__a, __m256i __b) 895{ 896 __builtin_ia32_movntdq256((__v4di *)__a, (__v4di)__b); 897} 898 899static __inline void __attribute__((__always_inline__, __nodebug__)) 900_mm256_stream_pd(double *__a, __m256d __b) 901{ 902 __builtin_ia32_movntpd256(__a, (__v4df)__b); 903} 904 905static __inline void __attribute__((__always_inline__, __nodebug__)) 906_mm256_stream_ps(float *__p, __m256 __a) 907{ 908 __builtin_ia32_movntps256(__p, (__v8sf)__a); 909} 910 911/* Create vectors */ 912static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 913_mm256_set_pd(double __a, double __b, double __c, double __d) 914{ 915 return (__m256d){ __d, __c, __b, __a }; 916} 917 918static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 919_mm256_set_ps(float __a, float __b, float __c, float __d, 920 float __e, float __f, float __g, float __h) 921{ 922 return (__m256){ __h, __g, __f, __e, __d, __c, __b, __a }; 923} 924 925static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 926_mm256_set_epi32(int __i0, int __i1, int __i2, int __i3, 927 int __i4, int __i5, int __i6, int __i7) 928{ 929 return (__m256i)(__v8si){ __i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0 }; 930} 931 932static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 933_mm256_set_epi16(short __w15, short __w14, short __w13, short __w12, 934 short __w11, short __w10, short __w09, short __w08, 935 short __w07, short __w06, short __w05, short __w04, 936 short __w03, short __w02, short __w01, short __w00) 937{ 938 return (__m256i)(__v16hi){ __w00, __w01, __w02, __w03, __w04, __w05, __w06, 939 __w07, __w08, __w09, __w10, __w11, __w12, __w13, __w14, __w15 }; 940} 941 942static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 943_mm256_set_epi8(char __b31, char __b30, char __b29, char __b28, 944 char __b27, char __b26, char __b25, char __b24, 945 char __b23, char __b22, char __b21, char __b20, 946 char __b19, char __b18, char __b17, char __b16, 947 char __b15, char __b14, char __b13, char __b12, 948 char __b11, char __b10, char __b09, char __b08, 949 char __b07, char __b06, char __b05, char __b04, 950 char __b03, char __b02, char __b01, char __b00) 951{ 952 return (__m256i)(__v32qi){ 953 __b00, __b01, __b02, __b03, __b04, __b05, __b06, __b07, 954 __b08, __b09, __b10, __b11, __b12, __b13, __b14, __b15, 955 __b16, __b17, __b18, __b19, __b20, __b21, __b22, __b23, 956 __b24, __b25, __b26, __b27, __b28, __b29, __b30, __b31 957 }; 958} 959 960static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 961_mm256_set_epi64x(long long __a, long long __b, long long __c, long long __d) 962{ 963 return (__m256i)(__v4di){ __d, __c, __b, __a }; 964} 965 966/* Create vectors with elements in reverse order */ 967static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 968_mm256_setr_pd(double __a, double __b, double __c, double __d) 969{ 970 return (__m256d){ __a, __b, __c, __d }; 971} 972 973static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 974_mm256_setr_ps(float __a, float __b, float __c, float __d, 975 float __e, float __f, float __g, float __h) 976{ 977 return (__m256){ __a, __b, __c, __d, __e, __f, __g, __h }; 978} 979 980static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 981_mm256_setr_epi32(int __i0, int __i1, int __i2, int __i3, 982 int __i4, int __i5, int __i6, int __i7) 983{ 984 return (__m256i)(__v8si){ __i0, __i1, __i2, __i3, __i4, __i5, __i6, __i7 }; 985} 986 987static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 988_mm256_setr_epi16(short __w15, short __w14, short __w13, short __w12, 989 short __w11, short __w10, short __w09, short __w08, 990 short __w07, short __w06, short __w05, short __w04, 991 short __w03, short __w02, short __w01, short __w00) 992{ 993 return (__m256i)(__v16hi){ __w15, __w14, __w13, __w12, __w11, __w10, __w09, 994 __w08, __w07, __w06, __w05, __w04, __w03, __w02, __w01, __w00 }; 995} 996 997static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 998_mm256_setr_epi8(char __b31, char __b30, char __b29, char __b28, 999 char __b27, char __b26, char __b25, char __b24, 1000 char __b23, char __b22, char __b21, char __b20, 1001 char __b19, char __b18, char __b17, char __b16, 1002 char __b15, char __b14, char __b13, char __b12, 1003 char __b11, char __b10, char __b09, char __b08, 1004 char __b07, char __b06, char __b05, char __b04, 1005 char __b03, char __b02, char __b01, char __b00) 1006{ 1007 return (__m256i)(__v32qi){ 1008 __b31, __b30, __b29, __b28, __b27, __b26, __b25, __b24, 1009 __b23, __b22, __b21, __b20, __b19, __b18, __b17, __b16, 1010 __b15, __b14, __b13, __b12, __b11, __b10, __b09, __b08, 1011 __b07, __b06, __b05, __b04, __b03, __b02, __b01, __b00 }; 1012} 1013 1014static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1015_mm256_setr_epi64x(long long __a, long long __b, long long __c, long long __d) 1016{ 1017 return (__m256i)(__v4di){ __a, __b, __c, __d }; 1018} 1019 1020/* Create vectors with repeated elements */ 1021static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1022_mm256_set1_pd(double __w) 1023{ 1024 return (__m256d){ __w, __w, __w, __w }; 1025} 1026 1027static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1028_mm256_set1_ps(float __w) 1029{ 1030 return (__m256){ __w, __w, __w, __w, __w, __w, __w, __w }; 1031} 1032 1033static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1034_mm256_set1_epi32(int __i) 1035{ 1036 return (__m256i)(__v8si){ __i, __i, __i, __i, __i, __i, __i, __i }; 1037} 1038 1039static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1040_mm256_set1_epi16(short __w) 1041{ 1042 return (__m256i)(__v16hi){ __w, __w, __w, __w, __w, __w, __w, __w, __w, __w, 1043 __w, __w, __w, __w, __w, __w }; 1044} 1045 1046static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1047_mm256_set1_epi8(char __b) 1048{ 1049 return (__m256i)(__v32qi){ __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, 1050 __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, __b, 1051 __b, __b, __b, __b, __b, __b, __b }; 1052} 1053 1054static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1055_mm256_set1_epi64x(long long __q) 1056{ 1057 return (__m256i)(__v4di){ __q, __q, __q, __q }; 1058} 1059 1060/* Create __zeroed vectors */ 1061static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1062_mm256_setzero_pd(void) 1063{ 1064 return (__m256d){ 0, 0, 0, 0 }; 1065} 1066 1067static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1068_mm256_setzero_ps(void) 1069{ 1070 return (__m256){ 0, 0, 0, 0, 0, 0, 0, 0 }; 1071} 1072 1073static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1074_mm256_setzero_si256(void) 1075{ 1076 return (__m256i){ 0LL, 0LL, 0LL, 0LL }; 1077} 1078 1079/* Cast between vector types */ 1080static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1081_mm256_castpd_ps(__m256d __in) 1082{ 1083 return (__m256)__in; 1084} 1085 1086static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1087_mm256_castpd_si256(__m256d __in) 1088{ 1089 return (__m256i)__in; 1090} 1091 1092static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1093_mm256_castps_pd(__m256 __in) 1094{ 1095 return (__m256d)__in; 1096} 1097 1098static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1099_mm256_castps_si256(__m256 __in) 1100{ 1101 return (__m256i)__in; 1102} 1103 1104static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1105_mm256_castsi256_ps(__m256i __in) 1106{ 1107 return (__m256)__in; 1108} 1109 1110static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1111_mm256_castsi256_pd(__m256i __in) 1112{ 1113 return (__m256d)__in; 1114} 1115 1116static __inline __m128d __attribute__((__always_inline__, __nodebug__)) 1117_mm256_castpd256_pd128(__m256d __in) 1118{ 1119 return __builtin_shufflevector(__in, __in, 0, 1); 1120} 1121 1122static __inline __m128 __attribute__((__always_inline__, __nodebug__)) 1123_mm256_castps256_ps128(__m256 __in) 1124{ 1125 return __builtin_shufflevector(__in, __in, 0, 1, 2, 3); 1126} 1127 1128static __inline __m128i __attribute__((__always_inline__, __nodebug__)) 1129_mm256_castsi256_si128(__m256i __in) 1130{ 1131 return __builtin_shufflevector(__in, __in, 0, 1); 1132} 1133 1134static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1135_mm256_castpd128_pd256(__m128d __in) 1136{ 1137 __m128d __zero = _mm_setzero_pd(); 1138 return __builtin_shufflevector(__in, __zero, 0, 1, 2, 2); 1139} 1140 1141static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1142_mm256_castps128_ps256(__m128 __in) 1143{ 1144 __m128 __zero = _mm_setzero_ps(); 1145 return __builtin_shufflevector(__in, __zero, 0, 1, 2, 3, 4, 4, 4, 4); 1146} 1147 1148static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1149_mm256_castsi128_si256(__m128i __in) 1150{ 1151 __m128i __zero = _mm_setzero_si128(); 1152 return __builtin_shufflevector(__in, __zero, 0, 1, 2, 2); 1153} 1154 1155/* SIMD load ops (unaligned) */ 1156static __inline __m256 __attribute__((__always_inline__, __nodebug__)) 1157_mm256_loadu2_m128(float const *__addr_hi, float const *__addr_lo) 1158{ 1159 struct __loadu_ps { 1160 __m128 __v; 1161 } __attribute__((__packed__, __may_alias__)); 1162 1163 __m256 __v256 = _mm256_castps128_ps256(((struct __loadu_ps*)__addr_lo)->__v); 1164 return _mm256_insertf128_ps(__v256, ((struct __loadu_ps*)__addr_hi)->__v, 1); 1165} 1166 1167static __inline __m256d __attribute__((__always_inline__, __nodebug__)) 1168_mm256_loadu2_m128d(double const *__addr_hi, double const *__addr_lo) 1169{ 1170 struct __loadu_pd { 1171 __m128d __v; 1172 } __attribute__((__packed__, __may_alias__)); 1173 1174 __m256d __v256 = _mm256_castpd128_pd256(((struct __loadu_pd*)__addr_lo)->__v); 1175 return _mm256_insertf128_pd(__v256, ((struct __loadu_pd*)__addr_hi)->__v, 1); 1176} 1177 1178static __inline __m256i __attribute__((__always_inline__, __nodebug__)) 1179_mm256_loadu2_m128i(__m128i const *__addr_hi, __m128i const *__addr_lo) 1180{ 1181 struct __loadu_si128 { 1182 __m128i __v; 1183 } __attribute__((packed, may_alias)); 1184 __m256i __v256 = _mm256_castsi128_si256( 1185 ((struct __loadu_si128*)__addr_lo)->__v); 1186 return _mm256_insertf128_si256(__v256, 1187 ((struct __loadu_si128*)__addr_hi)->__v, 1); 1188} 1189 1190/* SIMD store ops (unaligned) */ 1191static __inline void __attribute__((__always_inline__, __nodebug__)) 1192_mm256_storeu2_m128(float *__addr_hi, float *__addr_lo, __m256 __a) 1193{ 1194 __m128 __v128; 1195 1196 __v128 = _mm256_castps256_ps128(__a); 1197 __builtin_ia32_storeups(__addr_lo, __v128); 1198 __v128 = _mm256_extractf128_ps(__a, 1); 1199 __builtin_ia32_storeups(__addr_hi, __v128); 1200} 1201 1202static __inline void __attribute__((__always_inline__, __nodebug__)) 1203_mm256_storeu2_m128d(double *__addr_hi, double *__addr_lo, __m256d __a) 1204{ 1205 __m128d __v128; 1206 1207 __v128 = _mm256_castpd256_pd128(__a); 1208 __builtin_ia32_storeupd(__addr_lo, __v128); 1209 __v128 = _mm256_extractf128_pd(__a, 1); 1210 __builtin_ia32_storeupd(__addr_hi, __v128); 1211} 1212 1213static __inline void __attribute__((__always_inline__, __nodebug__)) 1214_mm256_storeu2_m128i(__m128i *__addr_hi, __m128i *__addr_lo, __m256i __a) 1215{ 1216 __m128i __v128; 1217 1218 __v128 = _mm256_castsi256_si128(__a); 1219 __builtin_ia32_storedqu((char *)__addr_lo, (__v16qi)__v128); 1220 __v128 = _mm256_extractf128_si256(__a, 1); 1221 __builtin_ia32_storedqu((char *)__addr_hi, (__v16qi)__v128); 1222} 1223