1 /*===---- avxintrin.h - AVX intrinsics -------------------------------------===
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:
10 * The above copyright notice and this permission notice shall be included in
11 * all copies or substantial portions of the Software.
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
21 *===-----------------------------------------------------------------------===
25 #error "Never use <avxintrin.h> directly; include <immintrin.h> instead."
28 typedef double __v4df __attribute__ ((__vector_size__ (32)));
29 typedef float __v8sf __attribute__ ((__vector_size__ (32)));
30 typedef long long __v4di __attribute__ ((__vector_size__ (32)));
31 typedef int __v8si __attribute__ ((__vector_size__ (32)));
32 typedef short __v16hi __attribute__ ((__vector_size__ (32)));
33 typedef char __v32qi __attribute__ ((__vector_size__ (32)));
35 typedef float __m256 __attribute__ ((__vector_size__ (32)));
36 typedef double __m256d __attribute__((__vector_size__(32)));
37 typedef long long __m256i __attribute__((__vector_size__(32)));
40 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
41 _mm256_add_pd(__m256d __a, __m256d __b)
46 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
47 _mm256_add_ps(__m256 __a, __m256 __b)
52 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
53 _mm256_sub_pd(__m256d __a, __m256d __b)
58 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
59 _mm256_sub_ps(__m256 __a, __m256 __b)
64 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
65 _mm256_addsub_pd(__m256d __a, __m256d __b)
67 return (__m256d)__builtin_ia32_addsubpd256((__v4df)__a, (__v4df)__b);
70 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
71 _mm256_addsub_ps(__m256 __a, __m256 __b)
73 return (__m256)__builtin_ia32_addsubps256((__v8sf)__a, (__v8sf)__b);
76 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
77 _mm256_div_pd(__m256d __a, __m256d __b)
82 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
83 _mm256_div_ps(__m256 __a, __m256 __b)
88 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
89 _mm256_max_pd(__m256d __a, __m256d __b)
91 return (__m256d)__builtin_ia32_maxpd256((__v4df)__a, (__v4df)__b);
94 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
95 _mm256_max_ps(__m256 __a, __m256 __b)
97 return (__m256)__builtin_ia32_maxps256((__v8sf)__a, (__v8sf)__b);
100 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
101 _mm256_min_pd(__m256d __a, __m256d __b)
103 return (__m256d)__builtin_ia32_minpd256((__v4df)__a, (__v4df)__b);
106 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
107 _mm256_min_ps(__m256 __a, __m256 __b)
109 return (__m256)__builtin_ia32_minps256((__v8sf)__a, (__v8sf)__b);
112 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
113 _mm256_mul_pd(__m256d __a, __m256d __b)
118 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
119 _mm256_mul_ps(__m256 __a, __m256 __b)
124 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
125 _mm256_sqrt_pd(__m256d __a)
127 return (__m256d)__builtin_ia32_sqrtpd256((__v4df)__a);
130 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
131 _mm256_sqrt_ps(__m256 __a)
133 return (__m256)__builtin_ia32_sqrtps256((__v8sf)__a);
136 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
137 _mm256_rsqrt_ps(__m256 __a)
139 return (__m256)__builtin_ia32_rsqrtps256((__v8sf)__a);
142 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
143 _mm256_rcp_ps(__m256 __a)
145 return (__m256)__builtin_ia32_rcpps256((__v8sf)__a);
148 #define _mm256_round_pd(V, M) __extension__ ({ \
150 (__m256d)__builtin_ia32_roundpd256((__v4df)__V, (M)); })
152 #define _mm256_round_ps(V, M) __extension__ ({ \
154 (__m256)__builtin_ia32_roundps256((__v8sf)__V, (M)); })
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)
162 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
163 _mm256_and_pd(__m256d __a, __m256d __b)
165 return (__m256d)((__v4di)__a & (__v4di)__b);
168 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
169 _mm256_and_ps(__m256 __a, __m256 __b)
171 return (__m256)((__v8si)__a & (__v8si)__b);
174 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
175 _mm256_andnot_pd(__m256d __a, __m256d __b)
177 return (__m256d)(~(__v4di)__a & (__v4di)__b);
180 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
181 _mm256_andnot_ps(__m256 __a, __m256 __b)
183 return (__m256)(~(__v8si)__a & (__v8si)__b);
186 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
187 _mm256_or_pd(__m256d __a, __m256d __b)
189 return (__m256d)((__v4di)__a | (__v4di)__b);
192 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
193 _mm256_or_ps(__m256 __a, __m256 __b)
195 return (__m256)((__v8si)__a | (__v8si)__b);
198 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
199 _mm256_xor_pd(__m256d __a, __m256d __b)
201 return (__m256d)((__v4di)__a ^ (__v4di)__b);
204 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
205 _mm256_xor_ps(__m256 __a, __m256 __b)
207 return (__m256)((__v8si)__a ^ (__v8si)__b);
210 /* Horizontal arithmetic */
211 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
212 _mm256_hadd_pd(__m256d __a, __m256d __b)
214 return (__m256d)__builtin_ia32_haddpd256((__v4df)__a, (__v4df)__b);
217 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
218 _mm256_hadd_ps(__m256 __a, __m256 __b)
220 return (__m256)__builtin_ia32_haddps256((__v8sf)__a, (__v8sf)__b);
223 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
224 _mm256_hsub_pd(__m256d __a, __m256d __b)
226 return (__m256d)__builtin_ia32_hsubpd256((__v4df)__a, (__v4df)__b);
229 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
230 _mm256_hsub_ps(__m256 __a, __m256 __b)
232 return (__m256)__builtin_ia32_hsubps256((__v8sf)__a, (__v8sf)__b);
235 /* Vector permutations */
236 static __inline __m128d __attribute__((__always_inline__, __nodebug__))
237 _mm_permutevar_pd(__m128d __a, __m128i __c)
239 return (__m128d)__builtin_ia32_vpermilvarpd((__v2df)__a, (__v2di)__c);
242 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
243 _mm256_permutevar_pd(__m256d __a, __m256i __c)
245 return (__m256d)__builtin_ia32_vpermilvarpd256((__v4df)__a, (__v4di)__c);
248 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
249 _mm_permutevar_ps(__m128 __a, __m128i __c)
251 return (__m128)__builtin_ia32_vpermilvarps((__v4sf)__a, (__v4si)__c);
254 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
255 _mm256_permutevar_ps(__m256 __a, __m256i __c)
257 return (__m256)__builtin_ia32_vpermilvarps256((__v8sf)__a,
261 #define _mm_permute_pd(A, C) __extension__ ({ \
263 (__m128d)__builtin_shufflevector((__v2df)__A, (__v2df) _mm_setzero_pd(), \
264 (C) & 0x1, ((C) & 0x2) >> 1); })
266 #define _mm256_permute_pd(A, C) __extension__ ({ \
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)); })
273 #define _mm_permute_ps(A, C) __extension__ ({ \
275 (__m128)__builtin_shufflevector((__v4sf)__A, (__v4sf) _mm_setzero_ps(), \
276 (C) & 0x3, ((C) & 0xc) >> 2, \
277 ((C) & 0x30) >> 4, ((C) & 0xc0) >> 6); })
279 #define _mm256_permute_ps(A, C) __extension__ ({ \
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)); })
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)); })
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)); })
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)); })
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)); })
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)); })
315 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
316 _mm256_blendv_pd(__m256d __a, __m256d __b, __m256d __c)
318 return (__m256d)__builtin_ia32_blendvpd256(
319 (__v4df)__a, (__v4df)__b, (__v4df)__c);
322 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
323 _mm256_blendv_ps(__m256 __a, __m256 __b, __m256 __c)
325 return (__m256)__builtin_ia32_blendvps256(
326 (__v8sf)__a, (__v8sf)__b, (__v8sf)__c);
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)); })
336 #define _mm256_shuffle_ps(a, b, mask) __extension__ ({ \
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); })
345 #define _mm256_shuffle_pd(a, b, mask) __extension__ ({ \
348 (__m256d)__builtin_shufflevector((__v4df)__a, (__v4df)__b, \
350 (((mask) & 0x2) >> 1) + 4, \
351 (((mask) & 0x4) >> 2) + 2, \
352 (((mask) & 0x8) >> 3) + 6); })
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) */
388 #define _mm_cmp_pd(a, b, c) __extension__ ({ \
391 (__m128d)__builtin_ia32_cmppd((__v2df)__a, (__v2df)__b, (c)); })
393 #define _mm_cmp_ps(a, b, c) __extension__ ({ \
396 (__m128)__builtin_ia32_cmpps((__v4sf)__a, (__v4sf)__b, (c)); })
398 #define _mm256_cmp_pd(a, b, c) __extension__ ({ \
401 (__m256d)__builtin_ia32_cmppd256((__v4df)__a, (__v4df)__b, (c)); })
403 #define _mm256_cmp_ps(a, b, c) __extension__ ({ \
406 (__m256)__builtin_ia32_cmpps256((__v8sf)__a, (__v8sf)__b, (c)); })
408 #define _mm_cmp_sd(a, b, c) __extension__ ({ \
411 (__m128d)__builtin_ia32_cmpsd((__v2df)__a, (__v2df)__b, (c)); })
413 #define _mm_cmp_ss(a, b, c) __extension__ ({ \
416 (__m128)__builtin_ia32_cmpss((__v4sf)__a, (__v4sf)__b, (c)); })
419 #define _mm256_extractf128_pd(A, O) __extension__ ({ \
421 (__m128d)__builtin_ia32_vextractf128_pd256((__v4df)__A, (O)); })
423 #define _mm256_extractf128_ps(A, O) __extension__ ({ \
425 (__m128)__builtin_ia32_vextractf128_ps256((__v8sf)__A, (O)); })
427 #define _mm256_extractf128_si256(A, O) __extension__ ({ \
429 (__m128i)__builtin_ia32_vextractf128_si256((__v8si)__A, (O)); })
431 static __inline int __attribute__((__always_inline__, __nodebug__))
432 _mm256_extract_epi32(__m256i __a, int const __imm)
434 __v8si __b = (__v8si)__a;
438 static __inline int __attribute__((__always_inline__, __nodebug__))
439 _mm256_extract_epi16(__m256i __a, int const __imm)
441 __v16hi __b = (__v16hi)__a;
445 static __inline int __attribute__((__always_inline__, __nodebug__))
446 _mm256_extract_epi8(__m256i __a, int const __imm)
448 __v32qi __b = (__v32qi)__a;
453 static __inline long long __attribute__((__always_inline__, __nodebug__))
454 _mm256_extract_epi64(__m256i __a, const int __imm)
456 __v4di __b = (__v4di)__a;
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)); })
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)); })
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)); })
477 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
478 _mm256_insert_epi32(__m256i __a, int __b, int const __imm)
480 __v8si __c = (__v8si)__a;
481 __c[__imm & 7] = __b;
485 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
486 _mm256_insert_epi16(__m256i __a, int __b, int const __imm)
488 __v16hi __c = (__v16hi)__a;
489 __c[__imm & 15] = __b;
493 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
494 _mm256_insert_epi8(__m256i __a, int __b, int const __imm)
496 __v32qi __c = (__v32qi)__a;
497 __c[__imm & 31] = __b;
502 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
503 _mm256_insert_epi64(__m256i __a, int __b, int const __imm)
505 __v4di __c = (__v4di)__a;
506 __c[__imm & 3] = __b;
512 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
513 _mm256_cvtepi32_pd(__m128i __a)
515 return (__m256d)__builtin_ia32_cvtdq2pd256((__v4si) __a);
518 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
519 _mm256_cvtepi32_ps(__m256i __a)
521 return (__m256)__builtin_ia32_cvtdq2ps256((__v8si) __a);
524 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
525 _mm256_cvtpd_ps(__m256d __a)
527 return (__m128)__builtin_ia32_cvtpd2ps256((__v4df) __a);
530 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
531 _mm256_cvtps_epi32(__m256 __a)
533 return (__m256i)__builtin_ia32_cvtps2dq256((__v8sf) __a);
536 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
537 _mm256_cvtps_pd(__m128 __a)
539 return (__m256d)__builtin_ia32_cvtps2pd256((__v4sf) __a);
542 static __inline __m128i __attribute__((__always_inline__, __nodebug__))
543 _mm256_cvttpd_epi32(__m256d __a)
545 return (__m128i)__builtin_ia32_cvttpd2dq256((__v4df) __a);
548 static __inline __m128i __attribute__((__always_inline__, __nodebug__))
549 _mm256_cvtpd_epi32(__m256d __a)
551 return (__m128i)__builtin_ia32_cvtpd2dq256((__v4df) __a);
554 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
555 _mm256_cvttps_epi32(__m256 __a)
557 return (__m256i)__builtin_ia32_cvttps2dq256((__v8sf) __a);
560 /* Vector replicate */
561 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
562 _mm256_movehdup_ps(__m256 __a)
564 return __builtin_shufflevector(__a, __a, 1, 1, 3, 3, 5, 5, 7, 7);
567 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
568 _mm256_moveldup_ps(__m256 __a)
570 return __builtin_shufflevector(__a, __a, 0, 0, 2, 2, 4, 4, 6, 6);
573 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
574 _mm256_movedup_pd(__m256d __a)
576 return __builtin_shufflevector(__a, __a, 0, 0, 2, 2);
579 /* Unpack and Interleave */
580 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
581 _mm256_unpackhi_pd(__m256d __a, __m256d __b)
583 return __builtin_shufflevector(__a, __b, 1, 5, 1+2, 5+2);
586 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
587 _mm256_unpacklo_pd(__m256d __a, __m256d __b)
589 return __builtin_shufflevector(__a, __b, 0, 4, 0+2, 4+2);
592 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
593 _mm256_unpackhi_ps(__m256 __a, __m256 __b)
595 return __builtin_shufflevector(__a, __b, 2, 10, 2+1, 10+1, 6, 14, 6+1, 14+1);
598 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
599 _mm256_unpacklo_ps(__m256 __a, __m256 __b)
601 return __builtin_shufflevector(__a, __b, 0, 8, 0+1, 8+1, 4, 12, 4+1, 12+1);
605 static __inline int __attribute__((__always_inline__, __nodebug__))
606 _mm_testz_pd(__m128d __a, __m128d __b)
608 return __builtin_ia32_vtestzpd((__v2df)__a, (__v2df)__b);
611 static __inline int __attribute__((__always_inline__, __nodebug__))
612 _mm_testc_pd(__m128d __a, __m128d __b)
614 return __builtin_ia32_vtestcpd((__v2df)__a, (__v2df)__b);
617 static __inline int __attribute__((__always_inline__, __nodebug__))
618 _mm_testnzc_pd(__m128d __a, __m128d __b)
620 return __builtin_ia32_vtestnzcpd((__v2df)__a, (__v2df)__b);
623 static __inline int __attribute__((__always_inline__, __nodebug__))
624 _mm_testz_ps(__m128 __a, __m128 __b)
626 return __builtin_ia32_vtestzps((__v4sf)__a, (__v4sf)__b);
629 static __inline int __attribute__((__always_inline__, __nodebug__))
630 _mm_testc_ps(__m128 __a, __m128 __b)
632 return __builtin_ia32_vtestcps((__v4sf)__a, (__v4sf)__b);
635 static __inline int __attribute__((__always_inline__, __nodebug__))
636 _mm_testnzc_ps(__m128 __a, __m128 __b)
638 return __builtin_ia32_vtestnzcps((__v4sf)__a, (__v4sf)__b);
641 static __inline int __attribute__((__always_inline__, __nodebug__))
642 _mm256_testz_pd(__m256d __a, __m256d __b)
644 return __builtin_ia32_vtestzpd256((__v4df)__a, (__v4df)__b);
647 static __inline int __attribute__((__always_inline__, __nodebug__))
648 _mm256_testc_pd(__m256d __a, __m256d __b)
650 return __builtin_ia32_vtestcpd256((__v4df)__a, (__v4df)__b);
653 static __inline int __attribute__((__always_inline__, __nodebug__))
654 _mm256_testnzc_pd(__m256d __a, __m256d __b)
656 return __builtin_ia32_vtestnzcpd256((__v4df)__a, (__v4df)__b);
659 static __inline int __attribute__((__always_inline__, __nodebug__))
660 _mm256_testz_ps(__m256 __a, __m256 __b)
662 return __builtin_ia32_vtestzps256((__v8sf)__a, (__v8sf)__b);
665 static __inline int __attribute__((__always_inline__, __nodebug__))
666 _mm256_testc_ps(__m256 __a, __m256 __b)
668 return __builtin_ia32_vtestcps256((__v8sf)__a, (__v8sf)__b);
671 static __inline int __attribute__((__always_inline__, __nodebug__))
672 _mm256_testnzc_ps(__m256 __a, __m256 __b)
674 return __builtin_ia32_vtestnzcps256((__v8sf)__a, (__v8sf)__b);
677 static __inline int __attribute__((__always_inline__, __nodebug__))
678 _mm256_testz_si256(__m256i __a, __m256i __b)
680 return __builtin_ia32_ptestz256((__v4di)__a, (__v4di)__b);
683 static __inline int __attribute__((__always_inline__, __nodebug__))
684 _mm256_testc_si256(__m256i __a, __m256i __b)
686 return __builtin_ia32_ptestc256((__v4di)__a, (__v4di)__b);
689 static __inline int __attribute__((__always_inline__, __nodebug__))
690 _mm256_testnzc_si256(__m256i __a, __m256i __b)
692 return __builtin_ia32_ptestnzc256((__v4di)__a, (__v4di)__b);
695 /* Vector extract sign mask */
696 static __inline int __attribute__((__always_inline__, __nodebug__))
697 _mm256_movemask_pd(__m256d __a)
699 return __builtin_ia32_movmskpd256((__v4df)__a);
702 static __inline int __attribute__((__always_inline__, __nodebug__))
703 _mm256_movemask_ps(__m256 __a)
705 return __builtin_ia32_movmskps256((__v8sf)__a);
709 static __inline void __attribute__((__always_inline__, __nodebug__))
712 __builtin_ia32_vzeroall();
715 static __inline void __attribute__((__always_inline__, __nodebug__))
716 _mm256_zeroupper(void)
718 __builtin_ia32_vzeroupper();
721 /* Vector load with broadcast */
722 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
723 _mm_broadcast_ss(float const *__a)
725 return (__m128)__builtin_ia32_vbroadcastss(__a);
728 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
729 _mm256_broadcast_sd(double const *__a)
731 return (__m256d)__builtin_ia32_vbroadcastsd256(__a);
734 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
735 _mm256_broadcast_ss(float const *__a)
737 return (__m256)__builtin_ia32_vbroadcastss256(__a);
740 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
741 _mm256_broadcast_pd(__m128d const *__a)
743 return (__m256d)__builtin_ia32_vbroadcastf128_pd256(__a);
746 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
747 _mm256_broadcast_ps(__m128 const *__a)
749 return (__m256)__builtin_ia32_vbroadcastf128_ps256(__a);
753 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
754 _mm256_load_pd(double const *__p)
756 return *(__m256d *)__p;
759 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
760 _mm256_load_ps(float const *__p)
762 return *(__m256 *)__p;
765 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
766 _mm256_loadu_pd(double const *__p)
770 } __attribute__((packed, may_alias));
771 return ((struct __loadu_pd*)__p)->__v;
774 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
775 _mm256_loadu_ps(float const *__p)
779 } __attribute__((packed, may_alias));
780 return ((struct __loadu_ps*)__p)->__v;
783 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
784 _mm256_load_si256(__m256i const *__p)
789 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
790 _mm256_loadu_si256(__m256i const *__p)
792 struct __loadu_si256 {
794 } __attribute__((packed, may_alias));
795 return ((struct __loadu_si256*)__p)->__v;
798 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
799 _mm256_lddqu_si256(__m256i const *__p)
801 return (__m256i)__builtin_ia32_lddqu256((char const *)__p);
805 static __inline void __attribute__((__always_inline__, __nodebug__))
806 _mm256_store_pd(double *__p, __m256d __a)
808 *(__m256d *)__p = __a;
811 static __inline void __attribute__((__always_inline__, __nodebug__))
812 _mm256_store_ps(float *__p, __m256 __a)
814 *(__m256 *)__p = __a;
817 static __inline void __attribute__((__always_inline__, __nodebug__))
818 _mm256_storeu_pd(double *__p, __m256d __a)
820 __builtin_ia32_storeupd256(__p, (__v4df)__a);
823 static __inline void __attribute__((__always_inline__, __nodebug__))
824 _mm256_storeu_ps(float *__p, __m256 __a)
826 __builtin_ia32_storeups256(__p, (__v8sf)__a);
829 static __inline void __attribute__((__always_inline__, __nodebug__))
830 _mm256_store_si256(__m256i *__p, __m256i __a)
835 static __inline void __attribute__((__always_inline__, __nodebug__))
836 _mm256_storeu_si256(__m256i *__p, __m256i __a)
838 __builtin_ia32_storedqu256((char *)__p, (__v32qi)__a);
841 /* Conditional load ops */
842 static __inline __m128d __attribute__((__always_inline__, __nodebug__))
843 _mm_maskload_pd(double const *__p, __m128d __m)
845 return (__m128d)__builtin_ia32_maskloadpd((const __v2df *)__p, (__v2df)__m);
848 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
849 _mm256_maskload_pd(double const *__p, __m256d __m)
851 return (__m256d)__builtin_ia32_maskloadpd256((const __v4df *)__p,
855 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
856 _mm_maskload_ps(float const *__p, __m128 __m)
858 return (__m128)__builtin_ia32_maskloadps((const __v4sf *)__p, (__v4sf)__m);
861 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
862 _mm256_maskload_ps(float const *__p, __m256 __m)
864 return (__m256)__builtin_ia32_maskloadps256((const __v8sf *)__p, (__v8sf)__m);
867 /* Conditional store ops */
868 static __inline void __attribute__((__always_inline__, __nodebug__))
869 _mm256_maskstore_ps(float *__p, __m256 __m, __m256 __a)
871 __builtin_ia32_maskstoreps256((__v8sf *)__p, (__v8sf)__m, (__v8sf)__a);
874 static __inline void __attribute__((__always_inline__, __nodebug__))
875 _mm_maskstore_pd(double *__p, __m128d __m, __m128d __a)
877 __builtin_ia32_maskstorepd((__v2df *)__p, (__v2df)__m, (__v2df)__a);
880 static __inline void __attribute__((__always_inline__, __nodebug__))
881 _mm256_maskstore_pd(double *__p, __m256d __m, __m256d __a)
883 __builtin_ia32_maskstorepd256((__v4df *)__p, (__v4df)__m, (__v4df)__a);
886 static __inline void __attribute__((__always_inline__, __nodebug__))
887 _mm_maskstore_ps(float *__p, __m128 __m, __m128 __a)
889 __builtin_ia32_maskstoreps((__v4sf *)__p, (__v4sf)__m, (__v4sf)__a);
892 /* Cacheability support ops */
893 static __inline void __attribute__((__always_inline__, __nodebug__))
894 _mm256_stream_si256(__m256i *__a, __m256i __b)
896 __builtin_ia32_movntdq256((__v4di *)__a, (__v4di)__b);
899 static __inline void __attribute__((__always_inline__, __nodebug__))
900 _mm256_stream_pd(double *__a, __m256d __b)
902 __builtin_ia32_movntpd256(__a, (__v4df)__b);
905 static __inline void __attribute__((__always_inline__, __nodebug__))
906 _mm256_stream_ps(float *__p, __m256 __a)
908 __builtin_ia32_movntps256(__p, (__v8sf)__a);
912 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
913 _mm256_set_pd(double __a, double __b, double __c, double __d)
915 return (__m256d){ __d, __c, __b, __a };
918 static __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)
922 return (__m256){ __h, __g, __f, __e, __d, __c, __b, __a };
925 static __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)
929 return (__m256i)(__v8si){ __i7, __i6, __i5, __i4, __i3, __i2, __i1, __i0 };
932 static __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)
938 return (__m256i)(__v16hi){ __w00, __w01, __w02, __w03, __w04, __w05, __w06,
939 __w07, __w08, __w09, __w10, __w11, __w12, __w13, __w14, __w15 };
942 static __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)
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
960 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
961 _mm256_set_epi64x(long long __a, long long __b, long long __c, long long __d)
963 return (__m256i)(__v4di){ __d, __c, __b, __a };
966 /* Create vectors with elements in reverse order */
967 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
968 _mm256_setr_pd(double __a, double __b, double __c, double __d)
970 return (__m256d){ __a, __b, __c, __d };
973 static __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)
977 return (__m256){ __a, __b, __c, __d, __e, __f, __g, __h };
980 static __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)
984 return (__m256i)(__v8si){ __i0, __i1, __i2, __i3, __i4, __i5, __i6, __i7 };
987 static __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)
993 return (__m256i)(__v16hi){ __w15, __w14, __w13, __w12, __w11, __w10, __w09,
994 __w08, __w07, __w06, __w05, __w04, __w03, __w02, __w01, __w00 };
997 static __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)
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 };
1014 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1015 _mm256_setr_epi64x(long long __a, long long __b, long long __c, long long __d)
1017 return (__m256i)(__v4di){ __a, __b, __c, __d };
1020 /* Create vectors with repeated elements */
1021 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
1022 _mm256_set1_pd(double __w)
1024 return (__m256d){ __w, __w, __w, __w };
1027 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
1028 _mm256_set1_ps(float __w)
1030 return (__m256){ __w, __w, __w, __w, __w, __w, __w, __w };
1033 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1034 _mm256_set1_epi32(int __i)
1036 return (__m256i)(__v8si){ __i, __i, __i, __i, __i, __i, __i, __i };
1039 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1040 _mm256_set1_epi16(short __w)
1042 return (__m256i)(__v16hi){ __w, __w, __w, __w, __w, __w, __w, __w, __w, __w,
1043 __w, __w, __w, __w, __w, __w };
1046 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1047 _mm256_set1_epi8(char __b)
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 };
1054 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1055 _mm256_set1_epi64x(long long __q)
1057 return (__m256i)(__v4di){ __q, __q, __q, __q };
1060 /* Create __zeroed vectors */
1061 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
1062 _mm256_setzero_pd(void)
1064 return (__m256d){ 0, 0, 0, 0 };
1067 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
1068 _mm256_setzero_ps(void)
1070 return (__m256){ 0, 0, 0, 0, 0, 0, 0, 0 };
1073 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1074 _mm256_setzero_si256(void)
1076 return (__m256i){ 0LL, 0LL, 0LL, 0LL };
1079 /* Cast between vector types */
1080 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
1081 _mm256_castpd_ps(__m256d __a)
1086 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1087 _mm256_castpd_si256(__m256d __a)
1089 return (__m256i)__a;
1092 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
1093 _mm256_castps_pd(__m256 __a)
1095 return (__m256d)__a;
1098 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1099 _mm256_castps_si256(__m256 __a)
1101 return (__m256i)__a;
1104 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
1105 _mm256_castsi256_ps(__m256i __a)
1110 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
1111 _mm256_castsi256_pd(__m256i __a)
1113 return (__m256d)__a;
1116 static __inline __m128d __attribute__((__always_inline__, __nodebug__))
1117 _mm256_castpd256_pd128(__m256d __a)
1119 return __builtin_shufflevector(__a, __a, 0, 1);
1122 static __inline __m128 __attribute__((__always_inline__, __nodebug__))
1123 _mm256_castps256_ps128(__m256 __a)
1125 return __builtin_shufflevector(__a, __a, 0, 1, 2, 3);
1128 static __inline __m128i __attribute__((__always_inline__, __nodebug__))
1129 _mm256_castsi256_si128(__m256i __a)
1131 return __builtin_shufflevector(__a, __a, 0, 1);
1134 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
1135 _mm256_castpd128_pd256(__m128d __a)
1137 __m128d __zero = _mm_setzero_pd();
1138 return __builtin_shufflevector(__a, __zero, 0, 1, 2, 2);
1141 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
1142 _mm256_castps128_ps256(__m128 __a)
1144 __m128 __zero = _mm_setzero_ps();
1145 return __builtin_shufflevector(__a, __zero, 0, 1, 2, 3, 4, 4, 4, 4);
1148 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1149 _mm256_castsi128_si256(__m128i __a)
1151 __m128i __zero = _mm_setzero_si128();
1152 return __builtin_shufflevector(__a, __zero, 0, 1, 2, 2);
1155 /* SIMD load ops (unaligned) */
1156 static __inline __m256 __attribute__((__always_inline__, __nodebug__))
1157 _mm256_loadu2_m128(float const *__addr_hi, float const *__addr_lo)
1161 } __attribute__((__packed__, __may_alias__));
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);
1167 static __inline __m256d __attribute__((__always_inline__, __nodebug__))
1168 _mm256_loadu2_m128d(double const *__addr_hi, double const *__addr_lo)
1172 } __attribute__((__packed__, __may_alias__));
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);
1178 static __inline __m256i __attribute__((__always_inline__, __nodebug__))
1179 _mm256_loadu2_m128i(__m128i const *__addr_hi, __m128i const *__addr_lo)
1181 struct __loadu_si128 {
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);
1190 /* SIMD store ops (unaligned) */
1191 static __inline void __attribute__((__always_inline__, __nodebug__))
1192 _mm256_storeu2_m128(float *__addr_hi, float *__addr_lo, __m256 __a)
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);
1202 static __inline void __attribute__((__always_inline__, __nodebug__))
1203 _mm256_storeu2_m128d(double *__addr_hi, double *__addr_lo, __m256d __a)
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);
1213 static __inline void __attribute__((__always_inline__, __nodebug__))
1214 _mm256_storeu2_m128i(__m128i *__addr_hi, __m128i *__addr_lo, __m256i __a)
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);