master
1/*===---- avx512dqintrin.h - AVX512DQ intrinsics ---------------------------===
2 *
3 * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
4 * See https://llvm.org/LICENSE.txt for license information.
5 * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
6 *
7 *===-----------------------------------------------------------------------===
8 */
9
10#ifndef __IMMINTRIN_H
11#error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead."
12#endif
13
14#ifndef __AVX512DQINTRIN_H
15#define __AVX512DQINTRIN_H
16
17/* Define the default attributes for the functions in this file. */
18#define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq,evex512"), __min_vector_width__(512)))
19#define __DEFAULT_FN_ATTRS \
20 __attribute__((__always_inline__, __nodebug__, \
21 __target__("avx512dq,no-evex512")))
22
23static __inline __mmask8 __DEFAULT_FN_ATTRS
24_knot_mask8(__mmask8 __M)
25{
26 return __builtin_ia32_knotqi(__M);
27}
28
29static __inline__ __mmask8 __DEFAULT_FN_ATTRS
30_kand_mask8(__mmask8 __A, __mmask8 __B)
31{
32 return (__mmask8)__builtin_ia32_kandqi((__mmask8)__A, (__mmask8)__B);
33}
34
35static __inline__ __mmask8 __DEFAULT_FN_ATTRS
36_kandn_mask8(__mmask8 __A, __mmask8 __B)
37{
38 return (__mmask8)__builtin_ia32_kandnqi((__mmask8)__A, (__mmask8)__B);
39}
40
41static __inline__ __mmask8 __DEFAULT_FN_ATTRS
42_kor_mask8(__mmask8 __A, __mmask8 __B)
43{
44 return (__mmask8)__builtin_ia32_korqi((__mmask8)__A, (__mmask8)__B);
45}
46
47static __inline__ __mmask8 __DEFAULT_FN_ATTRS
48_kxnor_mask8(__mmask8 __A, __mmask8 __B)
49{
50 return (__mmask8)__builtin_ia32_kxnorqi((__mmask8)__A, (__mmask8)__B);
51}
52
53static __inline__ __mmask8 __DEFAULT_FN_ATTRS
54_kxor_mask8(__mmask8 __A, __mmask8 __B)
55{
56 return (__mmask8)__builtin_ia32_kxorqi((__mmask8)__A, (__mmask8)__B);
57}
58
59static __inline__ unsigned char __DEFAULT_FN_ATTRS
60_kortestc_mask8_u8(__mmask8 __A, __mmask8 __B)
61{
62 return (unsigned char)__builtin_ia32_kortestcqi(__A, __B);
63}
64
65static __inline__ unsigned char __DEFAULT_FN_ATTRS
66_kortestz_mask8_u8(__mmask8 __A, __mmask8 __B)
67{
68 return (unsigned char)__builtin_ia32_kortestzqi(__A, __B);
69}
70
71static __inline__ unsigned char __DEFAULT_FN_ATTRS
72_kortest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) {
73 *__C = (unsigned char)__builtin_ia32_kortestcqi(__A, __B);
74 return (unsigned char)__builtin_ia32_kortestzqi(__A, __B);
75}
76
77static __inline__ unsigned char __DEFAULT_FN_ATTRS
78_ktestc_mask8_u8(__mmask8 __A, __mmask8 __B)
79{
80 return (unsigned char)__builtin_ia32_ktestcqi(__A, __B);
81}
82
83static __inline__ unsigned char __DEFAULT_FN_ATTRS
84_ktestz_mask8_u8(__mmask8 __A, __mmask8 __B)
85{
86 return (unsigned char)__builtin_ia32_ktestzqi(__A, __B);
87}
88
89static __inline__ unsigned char __DEFAULT_FN_ATTRS
90_ktest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C) {
91 *__C = (unsigned char)__builtin_ia32_ktestcqi(__A, __B);
92 return (unsigned char)__builtin_ia32_ktestzqi(__A, __B);
93}
94
95static __inline__ unsigned char __DEFAULT_FN_ATTRS
96_ktestc_mask16_u8(__mmask16 __A, __mmask16 __B)
97{
98 return (unsigned char)__builtin_ia32_ktestchi(__A, __B);
99}
100
101static __inline__ unsigned char __DEFAULT_FN_ATTRS
102_ktestz_mask16_u8(__mmask16 __A, __mmask16 __B)
103{
104 return (unsigned char)__builtin_ia32_ktestzhi(__A, __B);
105}
106
107static __inline__ unsigned char __DEFAULT_FN_ATTRS
108_ktest_mask16_u8(__mmask16 __A, __mmask16 __B, unsigned char *__C) {
109 *__C = (unsigned char)__builtin_ia32_ktestchi(__A, __B);
110 return (unsigned char)__builtin_ia32_ktestzhi(__A, __B);
111}
112
113static __inline__ __mmask8 __DEFAULT_FN_ATTRS
114_kadd_mask8(__mmask8 __A, __mmask8 __B)
115{
116 return (__mmask8)__builtin_ia32_kaddqi((__mmask8)__A, (__mmask8)__B);
117}
118
119static __inline__ __mmask16 __DEFAULT_FN_ATTRS
120_kadd_mask16(__mmask16 __A, __mmask16 __B)
121{
122 return (__mmask16)__builtin_ia32_kaddhi((__mmask16)__A, (__mmask16)__B);
123}
124
125#define _kshiftli_mask8(A, I) \
126 ((__mmask8)__builtin_ia32_kshiftliqi((__mmask8)(A), (unsigned int)(I)))
127
128#define _kshiftri_mask8(A, I) \
129 ((__mmask8)__builtin_ia32_kshiftriqi((__mmask8)(A), (unsigned int)(I)))
130
131static __inline__ unsigned int __DEFAULT_FN_ATTRS
132_cvtmask8_u32(__mmask8 __A) {
133 return (unsigned int)__builtin_ia32_kmovb((__mmask8)__A);
134}
135
136static __inline__ __mmask8 __DEFAULT_FN_ATTRS
137_cvtu32_mask8(unsigned int __A) {
138 return (__mmask8)__builtin_ia32_kmovb((__mmask8)__A);
139}
140
141static __inline__ __mmask8 __DEFAULT_FN_ATTRS
142_load_mask8(__mmask8 *__A) {
143 return (__mmask8)__builtin_ia32_kmovb(*(__mmask8 *)__A);
144}
145
146static __inline__ void __DEFAULT_FN_ATTRS
147_store_mask8(__mmask8 *__A, __mmask8 __B) {
148 *(__mmask8 *)__A = __builtin_ia32_kmovb((__mmask8)__B);
149}
150
151static __inline__ __m512i __DEFAULT_FN_ATTRS512
152_mm512_mullo_epi64 (__m512i __A, __m512i __B) {
153 return (__m512i) ((__v8du) __A * (__v8du) __B);
154}
155
156static __inline__ __m512i __DEFAULT_FN_ATTRS512
157_mm512_mask_mullo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B) {
158 return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
159 (__v8di)_mm512_mullo_epi64(__A, __B),
160 (__v8di)__W);
161}
162
163static __inline__ __m512i __DEFAULT_FN_ATTRS512
164_mm512_maskz_mullo_epi64(__mmask8 __U, __m512i __A, __m512i __B) {
165 return (__m512i)__builtin_ia32_selectq_512((__mmask8)__U,
166 (__v8di)_mm512_mullo_epi64(__A, __B),
167 (__v8di)_mm512_setzero_si512());
168}
169
170static __inline__ __m512d __DEFAULT_FN_ATTRS512
171_mm512_xor_pd(__m512d __A, __m512d __B) {
172 return (__m512d)((__v8du)__A ^ (__v8du)__B);
173}
174
175static __inline__ __m512d __DEFAULT_FN_ATTRS512
176_mm512_mask_xor_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
177 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
178 (__v8df)_mm512_xor_pd(__A, __B),
179 (__v8df)__W);
180}
181
182static __inline__ __m512d __DEFAULT_FN_ATTRS512
183_mm512_maskz_xor_pd(__mmask8 __U, __m512d __A, __m512d __B) {
184 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
185 (__v8df)_mm512_xor_pd(__A, __B),
186 (__v8df)_mm512_setzero_pd());
187}
188
189static __inline__ __m512 __DEFAULT_FN_ATTRS512
190_mm512_xor_ps (__m512 __A, __m512 __B) {
191 return (__m512)((__v16su)__A ^ (__v16su)__B);
192}
193
194static __inline__ __m512 __DEFAULT_FN_ATTRS512
195_mm512_mask_xor_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
196 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
197 (__v16sf)_mm512_xor_ps(__A, __B),
198 (__v16sf)__W);
199}
200
201static __inline__ __m512 __DEFAULT_FN_ATTRS512
202_mm512_maskz_xor_ps(__mmask16 __U, __m512 __A, __m512 __B) {
203 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
204 (__v16sf)_mm512_xor_ps(__A, __B),
205 (__v16sf)_mm512_setzero_ps());
206}
207
208static __inline__ __m512d __DEFAULT_FN_ATTRS512
209_mm512_or_pd(__m512d __A, __m512d __B) {
210 return (__m512d)((__v8du)__A | (__v8du)__B);
211}
212
213static __inline__ __m512d __DEFAULT_FN_ATTRS512
214_mm512_mask_or_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
215 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
216 (__v8df)_mm512_or_pd(__A, __B),
217 (__v8df)__W);
218}
219
220static __inline__ __m512d __DEFAULT_FN_ATTRS512
221_mm512_maskz_or_pd(__mmask8 __U, __m512d __A, __m512d __B) {
222 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
223 (__v8df)_mm512_or_pd(__A, __B),
224 (__v8df)_mm512_setzero_pd());
225}
226
227static __inline__ __m512 __DEFAULT_FN_ATTRS512
228_mm512_or_ps(__m512 __A, __m512 __B) {
229 return (__m512)((__v16su)__A | (__v16su)__B);
230}
231
232static __inline__ __m512 __DEFAULT_FN_ATTRS512
233_mm512_mask_or_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
234 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
235 (__v16sf)_mm512_or_ps(__A, __B),
236 (__v16sf)__W);
237}
238
239static __inline__ __m512 __DEFAULT_FN_ATTRS512
240_mm512_maskz_or_ps(__mmask16 __U, __m512 __A, __m512 __B) {
241 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
242 (__v16sf)_mm512_or_ps(__A, __B),
243 (__v16sf)_mm512_setzero_ps());
244}
245
246static __inline__ __m512d __DEFAULT_FN_ATTRS512
247_mm512_and_pd(__m512d __A, __m512d __B) {
248 return (__m512d)((__v8du)__A & (__v8du)__B);
249}
250
251static __inline__ __m512d __DEFAULT_FN_ATTRS512
252_mm512_mask_and_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
253 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
254 (__v8df)_mm512_and_pd(__A, __B),
255 (__v8df)__W);
256}
257
258static __inline__ __m512d __DEFAULT_FN_ATTRS512
259_mm512_maskz_and_pd(__mmask8 __U, __m512d __A, __m512d __B) {
260 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
261 (__v8df)_mm512_and_pd(__A, __B),
262 (__v8df)_mm512_setzero_pd());
263}
264
265static __inline__ __m512 __DEFAULT_FN_ATTRS512
266_mm512_and_ps(__m512 __A, __m512 __B) {
267 return (__m512)((__v16su)__A & (__v16su)__B);
268}
269
270static __inline__ __m512 __DEFAULT_FN_ATTRS512
271_mm512_mask_and_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
272 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
273 (__v16sf)_mm512_and_ps(__A, __B),
274 (__v16sf)__W);
275}
276
277static __inline__ __m512 __DEFAULT_FN_ATTRS512
278_mm512_maskz_and_ps(__mmask16 __U, __m512 __A, __m512 __B) {
279 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
280 (__v16sf)_mm512_and_ps(__A, __B),
281 (__v16sf)_mm512_setzero_ps());
282}
283
284static __inline__ __m512d __DEFAULT_FN_ATTRS512
285_mm512_andnot_pd(__m512d __A, __m512d __B) {
286 return (__m512d)(~(__v8du)__A & (__v8du)__B);
287}
288
289static __inline__ __m512d __DEFAULT_FN_ATTRS512
290_mm512_mask_andnot_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B) {
291 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
292 (__v8df)_mm512_andnot_pd(__A, __B),
293 (__v8df)__W);
294}
295
296static __inline__ __m512d __DEFAULT_FN_ATTRS512
297_mm512_maskz_andnot_pd(__mmask8 __U, __m512d __A, __m512d __B) {
298 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
299 (__v8df)_mm512_andnot_pd(__A, __B),
300 (__v8df)_mm512_setzero_pd());
301}
302
303static __inline__ __m512 __DEFAULT_FN_ATTRS512
304_mm512_andnot_ps(__m512 __A, __m512 __B) {
305 return (__m512)(~(__v16su)__A & (__v16su)__B);
306}
307
308static __inline__ __m512 __DEFAULT_FN_ATTRS512
309_mm512_mask_andnot_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B) {
310 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
311 (__v16sf)_mm512_andnot_ps(__A, __B),
312 (__v16sf)__W);
313}
314
315static __inline__ __m512 __DEFAULT_FN_ATTRS512
316_mm512_maskz_andnot_ps(__mmask16 __U, __m512 __A, __m512 __B) {
317 return (__m512)__builtin_ia32_selectps_512((__mmask16)__U,
318 (__v16sf)_mm512_andnot_ps(__A, __B),
319 (__v16sf)_mm512_setzero_ps());
320}
321
322static __inline__ __m512i __DEFAULT_FN_ATTRS512
323_mm512_cvtpd_epi64 (__m512d __A) {
324 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
325 (__v8di) _mm512_setzero_si512(),
326 (__mmask8) -1,
327 _MM_FROUND_CUR_DIRECTION);
328}
329
330static __inline__ __m512i __DEFAULT_FN_ATTRS512
331_mm512_mask_cvtpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) {
332 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
333 (__v8di) __W,
334 (__mmask8) __U,
335 _MM_FROUND_CUR_DIRECTION);
336}
337
338static __inline__ __m512i __DEFAULT_FN_ATTRS512
339_mm512_maskz_cvtpd_epi64 (__mmask8 __U, __m512d __A) {
340 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
341 (__v8di) _mm512_setzero_si512(),
342 (__mmask8) __U,
343 _MM_FROUND_CUR_DIRECTION);
344}
345
346#define _mm512_cvt_roundpd_epi64(A, R) \
347 ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
348 (__v8di)_mm512_setzero_si512(), \
349 (__mmask8)-1, (int)(R)))
350
351#define _mm512_mask_cvt_roundpd_epi64(W, U, A, R) \
352 ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
353 (__v8di)(__m512i)(W), \
354 (__mmask8)(U), (int)(R)))
355
356#define _mm512_maskz_cvt_roundpd_epi64(U, A, R) \
357 ((__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \
358 (__v8di)_mm512_setzero_si512(), \
359 (__mmask8)(U), (int)(R)))
360
361static __inline__ __m512i __DEFAULT_FN_ATTRS512
362_mm512_cvtpd_epu64 (__m512d __A) {
363 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
364 (__v8di) _mm512_setzero_si512(),
365 (__mmask8) -1,
366 _MM_FROUND_CUR_DIRECTION);
367}
368
369static __inline__ __m512i __DEFAULT_FN_ATTRS512
370_mm512_mask_cvtpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) {
371 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
372 (__v8di) __W,
373 (__mmask8) __U,
374 _MM_FROUND_CUR_DIRECTION);
375}
376
377static __inline__ __m512i __DEFAULT_FN_ATTRS512
378_mm512_maskz_cvtpd_epu64 (__mmask8 __U, __m512d __A) {
379 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
380 (__v8di) _mm512_setzero_si512(),
381 (__mmask8) __U,
382 _MM_FROUND_CUR_DIRECTION);
383}
384
385#define _mm512_cvt_roundpd_epu64(A, R) \
386 ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
387 (__v8di)_mm512_setzero_si512(), \
388 (__mmask8)-1, (int)(R)))
389
390#define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) \
391 ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
392 (__v8di)(__m512i)(W), \
393 (__mmask8)(U), (int)(R)))
394
395#define _mm512_maskz_cvt_roundpd_epu64(U, A, R) \
396 ((__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \
397 (__v8di)_mm512_setzero_si512(), \
398 (__mmask8)(U), (int)(R)))
399
400static __inline__ __m512i __DEFAULT_FN_ATTRS512
401_mm512_cvtps_epi64 (__m256 __A) {
402 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
403 (__v8di) _mm512_setzero_si512(),
404 (__mmask8) -1,
405 _MM_FROUND_CUR_DIRECTION);
406}
407
408static __inline__ __m512i __DEFAULT_FN_ATTRS512
409_mm512_mask_cvtps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) {
410 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
411 (__v8di) __W,
412 (__mmask8) __U,
413 _MM_FROUND_CUR_DIRECTION);
414}
415
416static __inline__ __m512i __DEFAULT_FN_ATTRS512
417_mm512_maskz_cvtps_epi64 (__mmask8 __U, __m256 __A) {
418 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
419 (__v8di) _mm512_setzero_si512(),
420 (__mmask8) __U,
421 _MM_FROUND_CUR_DIRECTION);
422}
423
424#define _mm512_cvt_roundps_epi64(A, R) \
425 ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
426 (__v8di)_mm512_setzero_si512(), \
427 (__mmask8)-1, (int)(R)))
428
429#define _mm512_mask_cvt_roundps_epi64(W, U, A, R) \
430 ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
431 (__v8di)(__m512i)(W), \
432 (__mmask8)(U), (int)(R)))
433
434#define _mm512_maskz_cvt_roundps_epi64(U, A, R) \
435 ((__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \
436 (__v8di)_mm512_setzero_si512(), \
437 (__mmask8)(U), (int)(R)))
438
439static __inline__ __m512i __DEFAULT_FN_ATTRS512
440_mm512_cvtps_epu64 (__m256 __A) {
441 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
442 (__v8di) _mm512_setzero_si512(),
443 (__mmask8) -1,
444 _MM_FROUND_CUR_DIRECTION);
445}
446
447static __inline__ __m512i __DEFAULT_FN_ATTRS512
448_mm512_mask_cvtps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) {
449 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
450 (__v8di) __W,
451 (__mmask8) __U,
452 _MM_FROUND_CUR_DIRECTION);
453}
454
455static __inline__ __m512i __DEFAULT_FN_ATTRS512
456_mm512_maskz_cvtps_epu64 (__mmask8 __U, __m256 __A) {
457 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
458 (__v8di) _mm512_setzero_si512(),
459 (__mmask8) __U,
460 _MM_FROUND_CUR_DIRECTION);
461}
462
463#define _mm512_cvt_roundps_epu64(A, R) \
464 ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
465 (__v8di)_mm512_setzero_si512(), \
466 (__mmask8)-1, (int)(R)))
467
468#define _mm512_mask_cvt_roundps_epu64(W, U, A, R) \
469 ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
470 (__v8di)(__m512i)(W), \
471 (__mmask8)(U), (int)(R)))
472
473#define _mm512_maskz_cvt_roundps_epu64(U, A, R) \
474 ((__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \
475 (__v8di)_mm512_setzero_si512(), \
476 (__mmask8)(U), (int)(R)))
477
478
479static __inline__ __m512d __DEFAULT_FN_ATTRS512
480_mm512_cvtepi64_pd (__m512i __A) {
481 return (__m512d)__builtin_convertvector((__v8di)__A, __v8df);
482}
483
484static __inline__ __m512d __DEFAULT_FN_ATTRS512
485_mm512_mask_cvtepi64_pd (__m512d __W, __mmask8 __U, __m512i __A) {
486 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
487 (__v8df)_mm512_cvtepi64_pd(__A),
488 (__v8df)__W);
489}
490
491static __inline__ __m512d __DEFAULT_FN_ATTRS512
492_mm512_maskz_cvtepi64_pd (__mmask8 __U, __m512i __A) {
493 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
494 (__v8df)_mm512_cvtepi64_pd(__A),
495 (__v8df)_mm512_setzero_pd());
496}
497
498#define _mm512_cvt_roundepi64_pd(A, R) \
499 ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
500 (__v8df)_mm512_setzero_pd(), \
501 (__mmask8)-1, (int)(R)))
502
503#define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) \
504 ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
505 (__v8df)(__m512d)(W), \
506 (__mmask8)(U), (int)(R)))
507
508#define _mm512_maskz_cvt_roundepi64_pd(U, A, R) \
509 ((__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \
510 (__v8df)_mm512_setzero_pd(), \
511 (__mmask8)(U), (int)(R)))
512
513static __inline__ __m256 __DEFAULT_FN_ATTRS512
514_mm512_cvtepi64_ps (__m512i __A) {
515 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
516 (__v8sf) _mm256_setzero_ps(),
517 (__mmask8) -1,
518 _MM_FROUND_CUR_DIRECTION);
519}
520
521static __inline__ __m256 __DEFAULT_FN_ATTRS512
522_mm512_mask_cvtepi64_ps (__m256 __W, __mmask8 __U, __m512i __A) {
523 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
524 (__v8sf) __W,
525 (__mmask8) __U,
526 _MM_FROUND_CUR_DIRECTION);
527}
528
529static __inline__ __m256 __DEFAULT_FN_ATTRS512
530_mm512_maskz_cvtepi64_ps (__mmask8 __U, __m512i __A) {
531 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
532 (__v8sf) _mm256_setzero_ps(),
533 (__mmask8) __U,
534 _MM_FROUND_CUR_DIRECTION);
535}
536
537#define _mm512_cvt_roundepi64_ps(A, R) \
538 ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
539 (__v8sf)_mm256_setzero_ps(), \
540 (__mmask8)-1, (int)(R)))
541
542#define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) \
543 ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
544 (__v8sf)(__m256)(W), (__mmask8)(U), \
545 (int)(R)))
546
547#define _mm512_maskz_cvt_roundepi64_ps(U, A, R) \
548 ((__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \
549 (__v8sf)_mm256_setzero_ps(), \
550 (__mmask8)(U), (int)(R)))
551
552
553static __inline__ __m512i __DEFAULT_FN_ATTRS512
554_mm512_cvttpd_epi64 (__m512d __A) {
555 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
556 (__v8di) _mm512_setzero_si512(),
557 (__mmask8) -1,
558 _MM_FROUND_CUR_DIRECTION);
559}
560
561static __inline__ __m512i __DEFAULT_FN_ATTRS512
562_mm512_mask_cvttpd_epi64 (__m512i __W, __mmask8 __U, __m512d __A) {
563 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
564 (__v8di) __W,
565 (__mmask8) __U,
566 _MM_FROUND_CUR_DIRECTION);
567}
568
569static __inline__ __m512i __DEFAULT_FN_ATTRS512
570_mm512_maskz_cvttpd_epi64 (__mmask8 __U, __m512d __A) {
571 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
572 (__v8di) _mm512_setzero_si512(),
573 (__mmask8) __U,
574 _MM_FROUND_CUR_DIRECTION);
575}
576
577#define _mm512_cvtt_roundpd_epi64(A, R) \
578 ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
579 (__v8di)_mm512_setzero_si512(), \
580 (__mmask8)-1, (int)(R)))
581
582#define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) \
583 ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
584 (__v8di)(__m512i)(W), \
585 (__mmask8)(U), (int)(R)))
586
587#define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) \
588 ((__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \
589 (__v8di)_mm512_setzero_si512(), \
590 (__mmask8)(U), (int)(R)))
591
592static __inline__ __m512i __DEFAULT_FN_ATTRS512
593_mm512_cvttpd_epu64 (__m512d __A) {
594 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
595 (__v8di) _mm512_setzero_si512(),
596 (__mmask8) -1,
597 _MM_FROUND_CUR_DIRECTION);
598}
599
600static __inline__ __m512i __DEFAULT_FN_ATTRS512
601_mm512_mask_cvttpd_epu64 (__m512i __W, __mmask8 __U, __m512d __A) {
602 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
603 (__v8di) __W,
604 (__mmask8) __U,
605 _MM_FROUND_CUR_DIRECTION);
606}
607
608static __inline__ __m512i __DEFAULT_FN_ATTRS512
609_mm512_maskz_cvttpd_epu64 (__mmask8 __U, __m512d __A) {
610 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
611 (__v8di) _mm512_setzero_si512(),
612 (__mmask8) __U,
613 _MM_FROUND_CUR_DIRECTION);
614}
615
616#define _mm512_cvtt_roundpd_epu64(A, R) \
617 ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
618 (__v8di)_mm512_setzero_si512(), \
619 (__mmask8)-1, (int)(R)))
620
621#define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) \
622 ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
623 (__v8di)(__m512i)(W), \
624 (__mmask8)(U), (int)(R)))
625
626#define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) \
627 ((__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \
628 (__v8di)_mm512_setzero_si512(), \
629 (__mmask8)(U), (int)(R)))
630
631static __inline__ __m512i __DEFAULT_FN_ATTRS512
632_mm512_cvttps_epi64 (__m256 __A) {
633 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
634 (__v8di) _mm512_setzero_si512(),
635 (__mmask8) -1,
636 _MM_FROUND_CUR_DIRECTION);
637}
638
639static __inline__ __m512i __DEFAULT_FN_ATTRS512
640_mm512_mask_cvttps_epi64 (__m512i __W, __mmask8 __U, __m256 __A) {
641 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
642 (__v8di) __W,
643 (__mmask8) __U,
644 _MM_FROUND_CUR_DIRECTION);
645}
646
647static __inline__ __m512i __DEFAULT_FN_ATTRS512
648_mm512_maskz_cvttps_epi64 (__mmask8 __U, __m256 __A) {
649 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
650 (__v8di) _mm512_setzero_si512(),
651 (__mmask8) __U,
652 _MM_FROUND_CUR_DIRECTION);
653}
654
655#define _mm512_cvtt_roundps_epi64(A, R) \
656 ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
657 (__v8di)_mm512_setzero_si512(), \
658 (__mmask8)-1, (int)(R)))
659
660#define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) \
661 ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
662 (__v8di)(__m512i)(W), \
663 (__mmask8)(U), (int)(R)))
664
665#define _mm512_maskz_cvtt_roundps_epi64(U, A, R) \
666 ((__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \
667 (__v8di)_mm512_setzero_si512(), \
668 (__mmask8)(U), (int)(R)))
669
670static __inline__ __m512i __DEFAULT_FN_ATTRS512
671_mm512_cvttps_epu64 (__m256 __A) {
672 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
673 (__v8di) _mm512_setzero_si512(),
674 (__mmask8) -1,
675 _MM_FROUND_CUR_DIRECTION);
676}
677
678static __inline__ __m512i __DEFAULT_FN_ATTRS512
679_mm512_mask_cvttps_epu64 (__m512i __W, __mmask8 __U, __m256 __A) {
680 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
681 (__v8di) __W,
682 (__mmask8) __U,
683 _MM_FROUND_CUR_DIRECTION);
684}
685
686static __inline__ __m512i __DEFAULT_FN_ATTRS512
687_mm512_maskz_cvttps_epu64 (__mmask8 __U, __m256 __A) {
688 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
689 (__v8di) _mm512_setzero_si512(),
690 (__mmask8) __U,
691 _MM_FROUND_CUR_DIRECTION);
692}
693
694#define _mm512_cvtt_roundps_epu64(A, R) \
695 ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
696 (__v8di)_mm512_setzero_si512(), \
697 (__mmask8)-1, (int)(R)))
698
699#define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) \
700 ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
701 (__v8di)(__m512i)(W), \
702 (__mmask8)(U), (int)(R)))
703
704#define _mm512_maskz_cvtt_roundps_epu64(U, A, R) \
705 ((__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \
706 (__v8di)_mm512_setzero_si512(), \
707 (__mmask8)(U), (int)(R)))
708
709static __inline__ __m512d __DEFAULT_FN_ATTRS512
710_mm512_cvtepu64_pd (__m512i __A) {
711 return (__m512d)__builtin_convertvector((__v8du)__A, __v8df);
712}
713
714static __inline__ __m512d __DEFAULT_FN_ATTRS512
715_mm512_mask_cvtepu64_pd (__m512d __W, __mmask8 __U, __m512i __A) {
716 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
717 (__v8df)_mm512_cvtepu64_pd(__A),
718 (__v8df)__W);
719}
720
721static __inline__ __m512d __DEFAULT_FN_ATTRS512
722_mm512_maskz_cvtepu64_pd (__mmask8 __U, __m512i __A) {
723 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__U,
724 (__v8df)_mm512_cvtepu64_pd(__A),
725 (__v8df)_mm512_setzero_pd());
726}
727
728#define _mm512_cvt_roundepu64_pd(A, R) \
729 ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
730 (__v8df)_mm512_setzero_pd(), \
731 (__mmask8)-1, (int)(R)))
732
733#define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) \
734 ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
735 (__v8df)(__m512d)(W), \
736 (__mmask8)(U), (int)(R)))
737
738
739#define _mm512_maskz_cvt_roundepu64_pd(U, A, R) \
740 ((__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \
741 (__v8df)_mm512_setzero_pd(), \
742 (__mmask8)(U), (int)(R)))
743
744
745static __inline__ __m256 __DEFAULT_FN_ATTRS512
746_mm512_cvtepu64_ps (__m512i __A) {
747 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
748 (__v8sf) _mm256_setzero_ps(),
749 (__mmask8) -1,
750 _MM_FROUND_CUR_DIRECTION);
751}
752
753static __inline__ __m256 __DEFAULT_FN_ATTRS512
754_mm512_mask_cvtepu64_ps (__m256 __W, __mmask8 __U, __m512i __A) {
755 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
756 (__v8sf) __W,
757 (__mmask8) __U,
758 _MM_FROUND_CUR_DIRECTION);
759}
760
761static __inline__ __m256 __DEFAULT_FN_ATTRS512
762_mm512_maskz_cvtepu64_ps (__mmask8 __U, __m512i __A) {
763 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
764 (__v8sf) _mm256_setzero_ps(),
765 (__mmask8) __U,
766 _MM_FROUND_CUR_DIRECTION);
767}
768
769#define _mm512_cvt_roundepu64_ps(A, R) \
770 ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
771 (__v8sf)_mm256_setzero_ps(), \
772 (__mmask8)-1, (int)(R)))
773
774#define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) \
775 ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
776 (__v8sf)(__m256)(W), (__mmask8)(U), \
777 (int)(R)))
778
779#define _mm512_maskz_cvt_roundepu64_ps(U, A, R) \
780 ((__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \
781 (__v8sf)_mm256_setzero_ps(), \
782 (__mmask8)(U), (int)(R)))
783
784#define _mm512_range_pd(A, B, C) \
785 ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
786 (__v8df)(__m512d)(B), (int)(C), \
787 (__v8df)_mm512_setzero_pd(), \
788 (__mmask8)-1, \
789 _MM_FROUND_CUR_DIRECTION))
790
791#define _mm512_mask_range_pd(W, U, A, B, C) \
792 ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
793 (__v8df)(__m512d)(B), (int)(C), \
794 (__v8df)(__m512d)(W), (__mmask8)(U), \
795 _MM_FROUND_CUR_DIRECTION))
796
797#define _mm512_maskz_range_pd(U, A, B, C) \
798 ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
799 (__v8df)(__m512d)(B), (int)(C), \
800 (__v8df)_mm512_setzero_pd(), \
801 (__mmask8)(U), \
802 _MM_FROUND_CUR_DIRECTION))
803
804#define _mm512_range_round_pd(A, B, C, R) \
805 ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
806 (__v8df)(__m512d)(B), (int)(C), \
807 (__v8df)_mm512_setzero_pd(), \
808 (__mmask8)-1, (int)(R)))
809
810#define _mm512_mask_range_round_pd(W, U, A, B, C, R) \
811 ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
812 (__v8df)(__m512d)(B), (int)(C), \
813 (__v8df)(__m512d)(W), (__mmask8)(U), \
814 (int)(R)))
815
816#define _mm512_maskz_range_round_pd(U, A, B, C, R) \
817 ((__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \
818 (__v8df)(__m512d)(B), (int)(C), \
819 (__v8df)_mm512_setzero_pd(), \
820 (__mmask8)(U), (int)(R)))
821
822#define _mm512_range_ps(A, B, C) \
823 ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
824 (__v16sf)(__m512)(B), (int)(C), \
825 (__v16sf)_mm512_setzero_ps(), \
826 (__mmask16)-1, \
827 _MM_FROUND_CUR_DIRECTION))
828
829#define _mm512_mask_range_ps(W, U, A, B, C) \
830 ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
831 (__v16sf)(__m512)(B), (int)(C), \
832 (__v16sf)(__m512)(W), (__mmask16)(U), \
833 _MM_FROUND_CUR_DIRECTION))
834
835#define _mm512_maskz_range_ps(U, A, B, C) \
836 ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
837 (__v16sf)(__m512)(B), (int)(C), \
838 (__v16sf)_mm512_setzero_ps(), \
839 (__mmask16)(U), \
840 _MM_FROUND_CUR_DIRECTION))
841
842#define _mm512_range_round_ps(A, B, C, R) \
843 ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
844 (__v16sf)(__m512)(B), (int)(C), \
845 (__v16sf)_mm512_setzero_ps(), \
846 (__mmask16)-1, (int)(R)))
847
848#define _mm512_mask_range_round_ps(W, U, A, B, C, R) \
849 ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
850 (__v16sf)(__m512)(B), (int)(C), \
851 (__v16sf)(__m512)(W), (__mmask16)(U), \
852 (int)(R)))
853
854#define _mm512_maskz_range_round_ps(U, A, B, C, R) \
855 ((__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \
856 (__v16sf)(__m512)(B), (int)(C), \
857 (__v16sf)_mm512_setzero_ps(), \
858 (__mmask16)(U), (int)(R)))
859
860#define _mm_range_round_ss(A, B, C, R) \
861 ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
862 (__v4sf)(__m128)(B), \
863 (__v4sf)_mm_setzero_ps(), \
864 (__mmask8) -1, (int)(C),\
865 (int)(R)))
866
867#define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION)
868
869#define _mm_mask_range_round_ss(W, U, A, B, C, R) \
870 ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
871 (__v4sf)(__m128)(B), \
872 (__v4sf)(__m128)(W),\
873 (__mmask8)(U), (int)(C),\
874 (int)(R)))
875
876#define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION)
877
878#define _mm_maskz_range_round_ss(U, A, B, C, R) \
879 ((__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \
880 (__v4sf)(__m128)(B), \
881 (__v4sf)_mm_setzero_ps(), \
882 (__mmask8)(U), (int)(C),\
883 (int)(R)))
884
885#define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
886
887#define _mm_range_round_sd(A, B, C, R) \
888 ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
889 (__v2df)(__m128d)(B), \
890 (__v2df)_mm_setzero_pd(), \
891 (__mmask8) -1, (int)(C),\
892 (int)(R)))
893
894#define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION)
895
896#define _mm_mask_range_round_sd(W, U, A, B, C, R) \
897 ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
898 (__v2df)(__m128d)(B), \
899 (__v2df)(__m128d)(W),\
900 (__mmask8)(U), (int)(C),\
901 (int)(R)))
902
903#define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
904
905#define _mm_maskz_range_round_sd(U, A, B, C, R) \
906 ((__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \
907 (__v2df)(__m128d)(B), \
908 (__v2df)_mm_setzero_pd(), \
909 (__mmask8)(U), (int)(C),\
910 (int)(R)))
911
912#define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION)
913
914#define _mm512_reduce_pd(A, B) \
915 ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
916 (__v8df)_mm512_setzero_pd(), \
917 (__mmask8)-1, \
918 _MM_FROUND_CUR_DIRECTION))
919
920#define _mm512_mask_reduce_pd(W, U, A, B) \
921 ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
922 (__v8df)(__m512d)(W), \
923 (__mmask8)(U), \
924 _MM_FROUND_CUR_DIRECTION))
925
926#define _mm512_maskz_reduce_pd(U, A, B) \
927 ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
928 (__v8df)_mm512_setzero_pd(), \
929 (__mmask8)(U), \
930 _MM_FROUND_CUR_DIRECTION))
931
932#define _mm512_reduce_ps(A, B) \
933 ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
934 (__v16sf)_mm512_setzero_ps(), \
935 (__mmask16)-1, \
936 _MM_FROUND_CUR_DIRECTION))
937
938#define _mm512_mask_reduce_ps(W, U, A, B) \
939 ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
940 (__v16sf)(__m512)(W), \
941 (__mmask16)(U), \
942 _MM_FROUND_CUR_DIRECTION))
943
944#define _mm512_maskz_reduce_ps(U, A, B) \
945 ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
946 (__v16sf)_mm512_setzero_ps(), \
947 (__mmask16)(U), \
948 _MM_FROUND_CUR_DIRECTION))
949
950#define _mm512_reduce_round_pd(A, B, R) \
951 ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
952 (__v8df)_mm512_setzero_pd(), \
953 (__mmask8)-1, (int)(R)))
954
955#define _mm512_mask_reduce_round_pd(W, U, A, B, R) \
956 ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
957 (__v8df)(__m512d)(W), \
958 (__mmask8)(U), (int)(R)))
959
960#define _mm512_maskz_reduce_round_pd(U, A, B, R) \
961 ((__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \
962 (__v8df)_mm512_setzero_pd(), \
963 (__mmask8)(U), (int)(R)))
964
965#define _mm512_reduce_round_ps(A, B, R) \
966 ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
967 (__v16sf)_mm512_setzero_ps(), \
968 (__mmask16)-1, (int)(R)))
969
970#define _mm512_mask_reduce_round_ps(W, U, A, B, R) \
971 ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
972 (__v16sf)(__m512)(W), \
973 (__mmask16)(U), (int)(R)))
974
975#define _mm512_maskz_reduce_round_ps(U, A, B, R) \
976 ((__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \
977 (__v16sf)_mm512_setzero_ps(), \
978 (__mmask16)(U), (int)(R)))
979
980#define _mm_reduce_ss(A, B, C) \
981 ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
982 (__v4sf)(__m128)(B), \
983 (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \
984 (int)(C), _MM_FROUND_CUR_DIRECTION))
985
986#define _mm_mask_reduce_ss(W, U, A, B, C) \
987 ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
988 (__v4sf)(__m128)(B), \
989 (__v4sf)(__m128)(W), (__mmask8)(U), \
990 (int)(C), _MM_FROUND_CUR_DIRECTION))
991
992#define _mm_maskz_reduce_ss(U, A, B, C) \
993 ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
994 (__v4sf)(__m128)(B), \
995 (__v4sf)_mm_setzero_ps(), \
996 (__mmask8)(U), (int)(C), \
997 _MM_FROUND_CUR_DIRECTION))
998
999#define _mm_reduce_round_ss(A, B, C, R) \
1000 ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1001 (__v4sf)(__m128)(B), \
1002 (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \
1003 (int)(C), (int)(R)))
1004
1005#define _mm_mask_reduce_round_ss(W, U, A, B, C, R) \
1006 ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1007 (__v4sf)(__m128)(B), \
1008 (__v4sf)(__m128)(W), (__mmask8)(U), \
1009 (int)(C), (int)(R)))
1010
1011#define _mm_maskz_reduce_round_ss(U, A, B, C, R) \
1012 ((__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \
1013 (__v4sf)(__m128)(B), \
1014 (__v4sf)_mm_setzero_ps(), \
1015 (__mmask8)(U), (int)(C), (int)(R)))
1016
1017#define _mm_reduce_sd(A, B, C) \
1018 ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1019 (__v2df)(__m128d)(B), \
1020 (__v2df)_mm_setzero_pd(), \
1021 (__mmask8)-1, (int)(C), \
1022 _MM_FROUND_CUR_DIRECTION))
1023
1024#define _mm_mask_reduce_sd(W, U, A, B, C) \
1025 ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1026 (__v2df)(__m128d)(B), \
1027 (__v2df)(__m128d)(W), (__mmask8)(U), \
1028 (int)(C), _MM_FROUND_CUR_DIRECTION))
1029
1030#define _mm_maskz_reduce_sd(U, A, B, C) \
1031 ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1032 (__v2df)(__m128d)(B), \
1033 (__v2df)_mm_setzero_pd(), \
1034 (__mmask8)(U), (int)(C), \
1035 _MM_FROUND_CUR_DIRECTION))
1036
1037#define _mm_reduce_round_sd(A, B, C, R) \
1038 ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1039 (__v2df)(__m128d)(B), \
1040 (__v2df)_mm_setzero_pd(), \
1041 (__mmask8)-1, (int)(C), (int)(R)))
1042
1043#define _mm_mask_reduce_round_sd(W, U, A, B, C, R) \
1044 ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1045 (__v2df)(__m128d)(B), \
1046 (__v2df)(__m128d)(W), (__mmask8)(U), \
1047 (int)(C), (int)(R)))
1048
1049#define _mm_maskz_reduce_round_sd(U, A, B, C, R) \
1050 ((__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \
1051 (__v2df)(__m128d)(B), \
1052 (__v2df)_mm_setzero_pd(), \
1053 (__mmask8)(U), (int)(C), (int)(R)))
1054
1055static __inline__ __mmask16 __DEFAULT_FN_ATTRS512
1056_mm512_movepi32_mask (__m512i __A)
1057{
1058 return (__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A);
1059}
1060
1061static __inline__ __m512i __DEFAULT_FN_ATTRS512
1062_mm512_movm_epi32 (__mmask16 __A)
1063{
1064 return (__m512i) __builtin_ia32_cvtmask2d512 (__A);
1065}
1066
1067static __inline__ __m512i __DEFAULT_FN_ATTRS512
1068_mm512_movm_epi64 (__mmask8 __A)
1069{
1070 return (__m512i) __builtin_ia32_cvtmask2q512 (__A);
1071}
1072
1073static __inline__ __mmask8 __DEFAULT_FN_ATTRS512
1074_mm512_movepi64_mask (__m512i __A)
1075{
1076 return (__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A);
1077}
1078
1079
1080static __inline__ __m512 __DEFAULT_FN_ATTRS512
1081_mm512_broadcast_f32x2 (__m128 __A)
1082{
1083 return (__m512)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A,
1084 0, 1, 0, 1, 0, 1, 0, 1,
1085 0, 1, 0, 1, 0, 1, 0, 1);
1086}
1087
1088static __inline__ __m512 __DEFAULT_FN_ATTRS512
1089_mm512_mask_broadcast_f32x2 (__m512 __O, __mmask16 __M, __m128 __A)
1090{
1091 return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1092 (__v16sf)_mm512_broadcast_f32x2(__A),
1093 (__v16sf)__O);
1094}
1095
1096static __inline__ __m512 __DEFAULT_FN_ATTRS512
1097_mm512_maskz_broadcast_f32x2 (__mmask16 __M, __m128 __A)
1098{
1099 return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1100 (__v16sf)_mm512_broadcast_f32x2(__A),
1101 (__v16sf)_mm512_setzero_ps());
1102}
1103
1104static __inline__ __m512 __DEFAULT_FN_ATTRS512
1105_mm512_broadcast_f32x8(__m256 __A)
1106{
1107 return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A,
1108 0, 1, 2, 3, 4, 5, 6, 7,
1109 0, 1, 2, 3, 4, 5, 6, 7);
1110}
1111
1112static __inline__ __m512 __DEFAULT_FN_ATTRS512
1113_mm512_mask_broadcast_f32x8(__m512 __O, __mmask16 __M, __m256 __A)
1114{
1115 return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1116 (__v16sf)_mm512_broadcast_f32x8(__A),
1117 (__v16sf)__O);
1118}
1119
1120static __inline__ __m512 __DEFAULT_FN_ATTRS512
1121_mm512_maskz_broadcast_f32x8(__mmask16 __M, __m256 __A)
1122{
1123 return (__m512)__builtin_ia32_selectps_512((__mmask16)__M,
1124 (__v16sf)_mm512_broadcast_f32x8(__A),
1125 (__v16sf)_mm512_setzero_ps());
1126}
1127
1128static __inline__ __m512d __DEFAULT_FN_ATTRS512
1129_mm512_broadcast_f64x2(__m128d __A)
1130{
1131 return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A,
1132 0, 1, 0, 1, 0, 1, 0, 1);
1133}
1134
1135static __inline__ __m512d __DEFAULT_FN_ATTRS512
1136_mm512_mask_broadcast_f64x2(__m512d __O, __mmask8 __M, __m128d __A)
1137{
1138 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M,
1139 (__v8df)_mm512_broadcast_f64x2(__A),
1140 (__v8df)__O);
1141}
1142
1143static __inline__ __m512d __DEFAULT_FN_ATTRS512
1144_mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A)
1145{
1146 return (__m512d)__builtin_ia32_selectpd_512((__mmask8)__M,
1147 (__v8df)_mm512_broadcast_f64x2(__A),
1148 (__v8df)_mm512_setzero_pd());
1149}
1150
1151static __inline__ __m512i __DEFAULT_FN_ATTRS512
1152_mm512_broadcast_i32x2 (__m128i __A)
1153{
1154 return (__m512i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
1155 0, 1, 0, 1, 0, 1, 0, 1,
1156 0, 1, 0, 1, 0, 1, 0, 1);
1157}
1158
1159static __inline__ __m512i __DEFAULT_FN_ATTRS512
1160_mm512_mask_broadcast_i32x2 (__m512i __O, __mmask16 __M, __m128i __A)
1161{
1162 return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1163 (__v16si)_mm512_broadcast_i32x2(__A),
1164 (__v16si)__O);
1165}
1166
1167static __inline__ __m512i __DEFAULT_FN_ATTRS512
1168_mm512_maskz_broadcast_i32x2 (__mmask16 __M, __m128i __A)
1169{
1170 return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1171 (__v16si)_mm512_broadcast_i32x2(__A),
1172 (__v16si)_mm512_setzero_si512());
1173}
1174
1175static __inline__ __m512i __DEFAULT_FN_ATTRS512
1176_mm512_broadcast_i32x8(__m256i __A)
1177{
1178 return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A,
1179 0, 1, 2, 3, 4, 5, 6, 7,
1180 0, 1, 2, 3, 4, 5, 6, 7);
1181}
1182
1183static __inline__ __m512i __DEFAULT_FN_ATTRS512
1184_mm512_mask_broadcast_i32x8(__m512i __O, __mmask16 __M, __m256i __A)
1185{
1186 return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1187 (__v16si)_mm512_broadcast_i32x8(__A),
1188 (__v16si)__O);
1189}
1190
1191static __inline__ __m512i __DEFAULT_FN_ATTRS512
1192_mm512_maskz_broadcast_i32x8(__mmask16 __M, __m256i __A)
1193{
1194 return (__m512i)__builtin_ia32_selectd_512((__mmask16)__M,
1195 (__v16si)_mm512_broadcast_i32x8(__A),
1196 (__v16si)_mm512_setzero_si512());
1197}
1198
1199static __inline__ __m512i __DEFAULT_FN_ATTRS512
1200_mm512_broadcast_i64x2(__m128i __A)
1201{
1202 return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A,
1203 0, 1, 0, 1, 0, 1, 0, 1);
1204}
1205
1206static __inline__ __m512i __DEFAULT_FN_ATTRS512
1207_mm512_mask_broadcast_i64x2(__m512i __O, __mmask8 __M, __m128i __A)
1208{
1209 return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M,
1210 (__v8di)_mm512_broadcast_i64x2(__A),
1211 (__v8di)__O);
1212}
1213
1214static __inline__ __m512i __DEFAULT_FN_ATTRS512
1215_mm512_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A)
1216{
1217 return (__m512i)__builtin_ia32_selectq_512((__mmask8)__M,
1218 (__v8di)_mm512_broadcast_i64x2(__A),
1219 (__v8di)_mm512_setzero_si512());
1220}
1221
1222#define _mm512_extractf32x8_ps(A, imm) \
1223 ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \
1224 (__v8sf)_mm256_undefined_ps(), \
1225 (__mmask8)-1))
1226
1227#define _mm512_mask_extractf32x8_ps(W, U, A, imm) \
1228 ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \
1229 (__v8sf)(__m256)(W), \
1230 (__mmask8)(U)))
1231
1232#define _mm512_maskz_extractf32x8_ps(U, A, imm) \
1233 ((__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \
1234 (__v8sf)_mm256_setzero_ps(), \
1235 (__mmask8)(U)))
1236
1237#define _mm512_extractf64x2_pd(A, imm) \
1238 ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \
1239 (int)(imm), \
1240 (__v2df)_mm_undefined_pd(), \
1241 (__mmask8)-1))
1242
1243#define _mm512_mask_extractf64x2_pd(W, U, A, imm) \
1244 ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \
1245 (int)(imm), \
1246 (__v2df)(__m128d)(W), \
1247 (__mmask8)(U)))
1248
1249#define _mm512_maskz_extractf64x2_pd(U, A, imm) \
1250 ((__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \
1251 (int)(imm), \
1252 (__v2df)_mm_setzero_pd(), \
1253 (__mmask8)(U)))
1254
1255#define _mm512_extracti32x8_epi32(A, imm) \
1256 ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \
1257 (__v8si)_mm256_undefined_si256(), \
1258 (__mmask8)-1))
1259
1260#define _mm512_mask_extracti32x8_epi32(W, U, A, imm) \
1261 ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \
1262 (__v8si)(__m256i)(W), \
1263 (__mmask8)(U)))
1264
1265#define _mm512_maskz_extracti32x8_epi32(U, A, imm) \
1266 ((__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \
1267 (__v8si)_mm256_setzero_si256(), \
1268 (__mmask8)(U)))
1269
1270#define _mm512_extracti64x2_epi64(A, imm) \
1271 ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \
1272 (int)(imm), \
1273 (__v2di)_mm_undefined_si128(), \
1274 (__mmask8)-1))
1275
1276#define _mm512_mask_extracti64x2_epi64(W, U, A, imm) \
1277 ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \
1278 (int)(imm), \
1279 (__v2di)(__m128i)(W), \
1280 (__mmask8)(U)))
1281
1282#define _mm512_maskz_extracti64x2_epi64(U, A, imm) \
1283 ((__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \
1284 (int)(imm), \
1285 (__v2di)_mm_setzero_si128(), \
1286 (__mmask8)(U)))
1287
1288#define _mm512_insertf32x8(A, B, imm) \
1289 ((__m512)__builtin_ia32_insertf32x8((__v16sf)(__m512)(A), \
1290 (__v8sf)(__m256)(B), (int)(imm)))
1291
1292#define _mm512_mask_insertf32x8(W, U, A, B, imm) \
1293 ((__m512)__builtin_ia32_selectps_512((__mmask16)(U), \
1294 (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \
1295 (__v16sf)(__m512)(W)))
1296
1297#define _mm512_maskz_insertf32x8(U, A, B, imm) \
1298 ((__m512)__builtin_ia32_selectps_512((__mmask16)(U), \
1299 (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \
1300 (__v16sf)_mm512_setzero_ps()))
1301
1302#define _mm512_insertf64x2(A, B, imm) \
1303 ((__m512d)__builtin_ia32_insertf64x2_512((__v8df)(__m512d)(A), \
1304 (__v2df)(__m128d)(B), (int)(imm)))
1305
1306#define _mm512_mask_insertf64x2(W, U, A, B, imm) \
1307 ((__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
1308 (__v8df)_mm512_insertf64x2((A), (B), (imm)), \
1309 (__v8df)(__m512d)(W)))
1310
1311#define _mm512_maskz_insertf64x2(U, A, B, imm) \
1312 ((__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \
1313 (__v8df)_mm512_insertf64x2((A), (B), (imm)), \
1314 (__v8df)_mm512_setzero_pd()))
1315
1316#define _mm512_inserti32x8(A, B, imm) \
1317 ((__m512i)__builtin_ia32_inserti32x8((__v16si)(__m512i)(A), \
1318 (__v8si)(__m256i)(B), (int)(imm)))
1319
1320#define _mm512_mask_inserti32x8(W, U, A, B, imm) \
1321 ((__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
1322 (__v16si)_mm512_inserti32x8((A), (B), (imm)), \
1323 (__v16si)(__m512i)(W)))
1324
1325#define _mm512_maskz_inserti32x8(U, A, B, imm) \
1326 ((__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \
1327 (__v16si)_mm512_inserti32x8((A), (B), (imm)), \
1328 (__v16si)_mm512_setzero_si512()))
1329
1330#define _mm512_inserti64x2(A, B, imm) \
1331 ((__m512i)__builtin_ia32_inserti64x2_512((__v8di)(__m512i)(A), \
1332 (__v2di)(__m128i)(B), (int)(imm)))
1333
1334#define _mm512_mask_inserti64x2(W, U, A, B, imm) \
1335 ((__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
1336 (__v8di)_mm512_inserti64x2((A), (B), (imm)), \
1337 (__v8di)(__m512i)(W)))
1338
1339#define _mm512_maskz_inserti64x2(U, A, B, imm) \
1340 ((__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \
1341 (__v8di)_mm512_inserti64x2((A), (B), (imm)), \
1342 (__v8di)_mm512_setzero_si512()))
1343
1344#define _mm512_mask_fpclass_ps_mask(U, A, imm) \
1345 ((__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \
1346 (int)(imm), (__mmask16)(U)))
1347
1348#define _mm512_fpclass_ps_mask(A, imm) \
1349 ((__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \
1350 (int)(imm), (__mmask16)-1))
1351
1352#define _mm512_mask_fpclass_pd_mask(U, A, imm) \
1353 ((__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \
1354 (__mmask8)(U)))
1355
1356#define _mm512_fpclass_pd_mask(A, imm) \
1357 ((__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \
1358 (__mmask8)-1))
1359
1360#define _mm_fpclass_sd_mask(A, imm) \
1361 ((__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \
1362 (__mmask8)-1))
1363
1364#define _mm_mask_fpclass_sd_mask(U, A, imm) \
1365 ((__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \
1366 (__mmask8)(U)))
1367
1368#define _mm_fpclass_ss_mask(A, imm) \
1369 ((__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \
1370 (__mmask8)-1))
1371
1372#define _mm_mask_fpclass_ss_mask(U, A, imm) \
1373 ((__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \
1374 (__mmask8)(U)))
1375
1376#undef __DEFAULT_FN_ATTRS512
1377#undef __DEFAULT_FN_ATTRS
1378
1379#endif