11 #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead." 14 #ifndef __AVX512DQINTRIN_H 15 #define __AVX512DQINTRIN_H 18 #define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"), __min_vector_width__(512))) 19 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"))) 24 return __builtin_ia32_knotqi(__M);
60 return (
unsigned char)__builtin_ia32_kortestcqi(__A, __B);
66 return (
unsigned char)__builtin_ia32_kortestzqi(__A, __B);
71 *__C = (
unsigned char)__builtin_ia32_kortestcqi(__A, __B);
72 return (
unsigned char)__builtin_ia32_kortestzqi(__A, __B);
78 return (
unsigned char)__builtin_ia32_ktestcqi(__A, __B);
84 return (
unsigned char)__builtin_ia32_ktestzqi(__A, __B);
89 *__C = (
unsigned char)__builtin_ia32_ktestcqi(__A, __B);
90 return (
unsigned char)__builtin_ia32_ktestzqi(__A, __B);
96 return (
unsigned char)__builtin_ia32_ktestchi(__A, __B);
102 return (
unsigned char)__builtin_ia32_ktestzhi(__A, __B);
107 *__C = (
unsigned char)__builtin_ia32_ktestchi(__A, __B);
108 return (
unsigned char)__builtin_ia32_ktestzhi(__A, __B);
123 #define _kshiftli_mask8(A, I) \ 124 (__mmask8)__builtin_ia32_kshiftliqi((__mmask8)(A), (unsigned int)(I)) 126 #define _kshiftri_mask8(A, I) \ 127 (__mmask8)__builtin_ia32_kshiftriqi((__mmask8)(A), (unsigned int)(I)) 131 return (
unsigned int)__builtin_ia32_kmovb((
__mmask8)__A);
151 return (__m512i) ((__v8du) __A * (__v8du) __B);
156 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__U,
163 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__U,
170 return (__m512d)((__v8du)__A ^ (__v8du)__B);
175 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
182 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
189 return (__m512)((__v16su)__A ^ (__v16su)__B);
194 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
201 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
208 return (__m512d)((__v8du)__A | (__v8du)__B);
213 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
220 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
227 return (__m512)((__v16su)__A | (__v16su)__B);
232 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
239 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
246 return (__m512d)((__v8du)__A & (__v8du)__B);
251 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
258 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
265 return (__m512)((__v16su)__A & (__v16su)__B);
270 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
277 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
284 return (__m512d)(~(__v8du)__A & (__v8du)__B);
289 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
296 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
303 return (__m512)(~(__v16su)__A & (__v16su)__B);
308 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
315 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
322 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
330 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
338 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
344 #define _mm512_cvt_roundpd_epi64(A, R) \ 345 (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ 346 (__v8di)_mm512_setzero_si512(), \ 347 (__mmask8)-1, (int)(R)) 349 #define _mm512_mask_cvt_roundpd_epi64(W, U, A, R) \ 350 (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ 351 (__v8di)(__m512i)(W), \ 352 (__mmask8)(U), (int)(R)) 354 #define _mm512_maskz_cvt_roundpd_epi64(U, A, R) \ 355 (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ 356 (__v8di)_mm512_setzero_si512(), \ 357 (__mmask8)(U), (int)(R)) 361 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
369 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
377 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
383 #define _mm512_cvt_roundpd_epu64(A, R) \ 384 (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ 385 (__v8di)_mm512_setzero_si512(), \ 386 (__mmask8)-1, (int)(R)) 388 #define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) \ 389 (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ 390 (__v8di)(__m512i)(W), \ 391 (__mmask8)(U), (int)(R)) 393 #define _mm512_maskz_cvt_roundpd_epu64(U, A, R) \ 394 (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ 395 (__v8di)_mm512_setzero_si512(), \ 396 (__mmask8)(U), (int)(R)) 400 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
408 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
416 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
422 #define _mm512_cvt_roundps_epi64(A, R) \ 423 (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ 424 (__v8di)_mm512_setzero_si512(), \ 425 (__mmask8)-1, (int)(R)) 427 #define _mm512_mask_cvt_roundps_epi64(W, U, A, R) \ 428 (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ 429 (__v8di)(__m512i)(W), \ 430 (__mmask8)(U), (int)(R)) 432 #define _mm512_maskz_cvt_roundps_epi64(U, A, R) \ 433 (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ 434 (__v8di)_mm512_setzero_si512(), \ 435 (__mmask8)(U), (int)(R)) 439 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
447 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
455 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
461 #define _mm512_cvt_roundps_epu64(A, R) \ 462 (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ 463 (__v8di)_mm512_setzero_si512(), \ 464 (__mmask8)-1, (int)(R)) 466 #define _mm512_mask_cvt_roundps_epu64(W, U, A, R) \ 467 (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ 468 (__v8di)(__m512i)(W), \ 469 (__mmask8)(U), (int)(R)) 471 #define _mm512_maskz_cvt_roundps_epu64(U, A, R) \ 472 (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ 473 (__v8di)_mm512_setzero_si512(), \ 474 (__mmask8)(U), (int)(R)) 479 return (__m512d)__builtin_convertvector((__v8di)__A, __v8df);
484 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
491 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
496 #define _mm512_cvt_roundepi64_pd(A, R) \ 497 (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ 498 (__v8df)_mm512_setzero_pd(), \ 499 (__mmask8)-1, (int)(R)) 501 #define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) \ 502 (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ 503 (__v8df)(__m512d)(W), \ 504 (__mmask8)(U), (int)(R)) 506 #define _mm512_maskz_cvt_roundepi64_pd(U, A, R) \ 507 (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ 508 (__v8df)_mm512_setzero_pd(), \ 509 (__mmask8)(U), (int)(R)) 513 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
521 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
529 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
535 #define _mm512_cvt_roundepi64_ps(A, R) \ 536 (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ 537 (__v8sf)_mm256_setzero_ps(), \ 538 (__mmask8)-1, (int)(R)) 540 #define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) \ 541 (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ 542 (__v8sf)(__m256)(W), (__mmask8)(U), \ 545 #define _mm512_maskz_cvt_roundepi64_ps(U, A, R) \ 546 (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ 547 (__v8sf)_mm256_setzero_ps(), \ 548 (__mmask8)(U), (int)(R)) 553 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
561 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
569 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
575 #define _mm512_cvtt_roundpd_epi64(A, R) \ 576 (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ 577 (__v8di)_mm512_setzero_si512(), \ 578 (__mmask8)-1, (int)(R)) 580 #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) \ 581 (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ 582 (__v8di)(__m512i)(W), \ 583 (__mmask8)(U), (int)(R)) 585 #define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) \ 586 (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ 587 (__v8di)_mm512_setzero_si512(), \ 588 (__mmask8)(U), (int)(R)) 592 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
600 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
608 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
614 #define _mm512_cvtt_roundpd_epu64(A, R) \ 615 (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ 616 (__v8di)_mm512_setzero_si512(), \ 617 (__mmask8)-1, (int)(R)) 619 #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) \ 620 (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ 621 (__v8di)(__m512i)(W), \ 622 (__mmask8)(U), (int)(R)) 624 #define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) \ 625 (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ 626 (__v8di)_mm512_setzero_si512(), \ 627 (__mmask8)(U), (int)(R)) 631 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
639 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
647 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
653 #define _mm512_cvtt_roundps_epi64(A, R) \ 654 (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ 655 (__v8di)_mm512_setzero_si512(), \ 656 (__mmask8)-1, (int)(R)) 658 #define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) \ 659 (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ 660 (__v8di)(__m512i)(W), \ 661 (__mmask8)(U), (int)(R)) 663 #define _mm512_maskz_cvtt_roundps_epi64(U, A, R) \ 664 (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ 665 (__v8di)_mm512_setzero_si512(), \ 666 (__mmask8)(U), (int)(R)) 670 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
678 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
686 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
692 #define _mm512_cvtt_roundps_epu64(A, R) \ 693 (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ 694 (__v8di)_mm512_setzero_si512(), \ 695 (__mmask8)-1, (int)(R)) 697 #define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) \ 698 (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ 699 (__v8di)(__m512i)(W), \ 700 (__mmask8)(U), (int)(R)) 702 #define _mm512_maskz_cvtt_roundps_epu64(U, A, R) \ 703 (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ 704 (__v8di)_mm512_setzero_si512(), \ 705 (__mmask8)(U), (int)(R)) 709 return (__m512d)__builtin_convertvector((__v8du)__A, __v8df);
714 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
721 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
726 #define _mm512_cvt_roundepu64_pd(A, R) \ 727 (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ 728 (__v8df)_mm512_setzero_pd(), \ 729 (__mmask8)-1, (int)(R)) 731 #define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) \ 732 (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ 733 (__v8df)(__m512d)(W), \ 734 (__mmask8)(U), (int)(R)) 737 #define _mm512_maskz_cvt_roundepu64_pd(U, A, R) \ 738 (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ 739 (__v8df)_mm512_setzero_pd(), \ 740 (__mmask8)(U), (int)(R)) 745 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
753 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
761 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
767 #define _mm512_cvt_roundepu64_ps(A, R) \ 768 (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ 769 (__v8sf)_mm256_setzero_ps(), \ 770 (__mmask8)-1, (int)(R)) 772 #define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) \ 773 (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ 774 (__v8sf)(__m256)(W), (__mmask8)(U), \ 777 #define _mm512_maskz_cvt_roundepu64_ps(U, A, R) \ 778 (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ 779 (__v8sf)_mm256_setzero_ps(), \ 780 (__mmask8)(U), (int)(R)) 782 #define _mm512_range_pd(A, B, C) \ 783 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 784 (__v8df)(__m512d)(B), (int)(C), \ 785 (__v8df)_mm512_setzero_pd(), \ 787 _MM_FROUND_CUR_DIRECTION) 789 #define _mm512_mask_range_pd(W, U, A, B, C) \ 790 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 791 (__v8df)(__m512d)(B), (int)(C), \ 792 (__v8df)(__m512d)(W), (__mmask8)(U), \ 793 _MM_FROUND_CUR_DIRECTION) 795 #define _mm512_maskz_range_pd(U, A, B, C) \ 796 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 797 (__v8df)(__m512d)(B), (int)(C), \ 798 (__v8df)_mm512_setzero_pd(), \ 800 _MM_FROUND_CUR_DIRECTION) 802 #define _mm512_range_round_pd(A, B, C, R) \ 803 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 804 (__v8df)(__m512d)(B), (int)(C), \ 805 (__v8df)_mm512_setzero_pd(), \ 806 (__mmask8)-1, (int)(R)) 808 #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ 809 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 810 (__v8df)(__m512d)(B), (int)(C), \ 811 (__v8df)(__m512d)(W), (__mmask8)(U), \ 814 #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ 815 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 816 (__v8df)(__m512d)(B), (int)(C), \ 817 (__v8df)_mm512_setzero_pd(), \ 818 (__mmask8)(U), (int)(R)) 820 #define _mm512_range_ps(A, B, C) \ 821 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 822 (__v16sf)(__m512)(B), (int)(C), \ 823 (__v16sf)_mm512_setzero_ps(), \ 825 _MM_FROUND_CUR_DIRECTION) 827 #define _mm512_mask_range_ps(W, U, A, B, C) \ 828 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 829 (__v16sf)(__m512)(B), (int)(C), \ 830 (__v16sf)(__m512)(W), (__mmask16)(U), \ 831 _MM_FROUND_CUR_DIRECTION) 833 #define _mm512_maskz_range_ps(U, A, B, C) \ 834 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 835 (__v16sf)(__m512)(B), (int)(C), \ 836 (__v16sf)_mm512_setzero_ps(), \ 838 _MM_FROUND_CUR_DIRECTION) 840 #define _mm512_range_round_ps(A, B, C, R) \ 841 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 842 (__v16sf)(__m512)(B), (int)(C), \ 843 (__v16sf)_mm512_setzero_ps(), \ 844 (__mmask16)-1, (int)(R)) 846 #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ 847 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 848 (__v16sf)(__m512)(B), (int)(C), \ 849 (__v16sf)(__m512)(W), (__mmask16)(U), \ 852 #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ 853 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 854 (__v16sf)(__m512)(B), (int)(C), \ 855 (__v16sf)_mm512_setzero_ps(), \ 856 (__mmask16)(U), (int)(R)) 858 #define _mm_range_round_ss(A, B, C, R) \ 859 (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ 860 (__v4sf)(__m128)(B), \ 861 (__v4sf)_mm_setzero_ps(), \ 862 (__mmask8) -1, (int)(C),\ 865 #define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION) 867 #define _mm_mask_range_round_ss(W, U, A, B, C, R) \ 868 (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ 869 (__v4sf)(__m128)(B), \ 870 (__v4sf)(__m128)(W),\ 871 (__mmask8)(U), (int)(C),\ 874 #define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION) 876 #define _mm_maskz_range_round_ss(U, A, B, C, R) \ 877 (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ 878 (__v4sf)(__m128)(B), \ 879 (__v4sf)_mm_setzero_ps(), \ 880 (__mmask8)(U), (int)(C),\ 883 #define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) 885 #define _mm_range_round_sd(A, B, C, R) \ 886 (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ 887 (__v2df)(__m128d)(B), \ 888 (__v2df)_mm_setzero_pd(), \ 889 (__mmask8) -1, (int)(C),\ 892 #define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION) 894 #define _mm_mask_range_round_sd(W, U, A, B, C, R) \ 895 (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ 896 (__v2df)(__m128d)(B), \ 897 (__v2df)(__m128d)(W),\ 898 (__mmask8)(U), (int)(C),\ 901 #define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION) 903 #define _mm_maskz_range_round_sd(U, A, B, C, R) \ 904 (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ 905 (__v2df)(__m128d)(B), \ 906 (__v2df)_mm_setzero_pd(), \ 907 (__mmask8)(U), (int)(C),\ 910 #define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) 912 #define _mm512_reduce_pd(A, B) \ 913 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 914 (__v8df)_mm512_setzero_pd(), \ 916 _MM_FROUND_CUR_DIRECTION) 918 #define _mm512_mask_reduce_pd(W, U, A, B) \ 919 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 920 (__v8df)(__m512d)(W), \ 922 _MM_FROUND_CUR_DIRECTION) 924 #define _mm512_maskz_reduce_pd(U, A, B) \ 925 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 926 (__v8df)_mm512_setzero_pd(), \ 928 _MM_FROUND_CUR_DIRECTION) 930 #define _mm512_reduce_ps(A, B) \ 931 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 932 (__v16sf)_mm512_setzero_ps(), \ 934 _MM_FROUND_CUR_DIRECTION) 936 #define _mm512_mask_reduce_ps(W, U, A, B) \ 937 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 938 (__v16sf)(__m512)(W), \ 940 _MM_FROUND_CUR_DIRECTION) 942 #define _mm512_maskz_reduce_ps(U, A, B) \ 943 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 944 (__v16sf)_mm512_setzero_ps(), \ 946 _MM_FROUND_CUR_DIRECTION) 948 #define _mm512_reduce_round_pd(A, B, R) \ 949 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 950 (__v8df)_mm512_setzero_pd(), \ 951 (__mmask8)-1, (int)(R)) 953 #define _mm512_mask_reduce_round_pd(W, U, A, B, R) \ 954 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 955 (__v8df)(__m512d)(W), \ 956 (__mmask8)(U), (int)(R)) 958 #define _mm512_maskz_reduce_round_pd(U, A, B, R) \ 959 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 960 (__v8df)_mm512_setzero_pd(), \ 961 (__mmask8)(U), (int)(R)) 963 #define _mm512_reduce_round_ps(A, B, R) \ 964 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 965 (__v16sf)_mm512_setzero_ps(), \ 966 (__mmask16)-1, (int)(R)) 968 #define _mm512_mask_reduce_round_ps(W, U, A, B, R) \ 969 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 970 (__v16sf)(__m512)(W), \ 971 (__mmask16)(U), (int)(R)) 973 #define _mm512_maskz_reduce_round_ps(U, A, B, R) \ 974 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 975 (__v16sf)_mm512_setzero_ps(), \ 976 (__mmask16)(U), (int)(R)) 978 #define _mm_reduce_ss(A, B, C) \ 979 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 980 (__v4sf)(__m128)(B), \ 981 (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ 982 (int)(C), _MM_FROUND_CUR_DIRECTION) 984 #define _mm_mask_reduce_ss(W, U, A, B, C) \ 985 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 986 (__v4sf)(__m128)(B), \ 987 (__v4sf)(__m128)(W), (__mmask8)(U), \ 988 (int)(C), _MM_FROUND_CUR_DIRECTION) 990 #define _mm_maskz_reduce_ss(U, A, B, C) \ 991 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 992 (__v4sf)(__m128)(B), \ 993 (__v4sf)_mm_setzero_ps(), \ 994 (__mmask8)(U), (int)(C), \ 995 _MM_FROUND_CUR_DIRECTION) 997 #define _mm_reduce_round_ss(A, B, C, R) \ 998 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 999 (__v4sf)(__m128)(B), \ 1000 (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ 1003 #define _mm_mask_reduce_round_ss(W, U, A, B, C, R) \ 1004 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1005 (__v4sf)(__m128)(B), \ 1006 (__v4sf)(__m128)(W), (__mmask8)(U), \ 1009 #define _mm_maskz_reduce_round_ss(U, A, B, C, R) \ 1010 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1011 (__v4sf)(__m128)(B), \ 1012 (__v4sf)_mm_setzero_ps(), \ 1013 (__mmask8)(U), (int)(C), (int)(R)) 1015 #define _mm_reduce_sd(A, B, C) \ 1016 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1017 (__v2df)(__m128d)(B), \ 1018 (__v2df)_mm_setzero_pd(), \ 1019 (__mmask8)-1, (int)(C), \ 1020 _MM_FROUND_CUR_DIRECTION) 1022 #define _mm_mask_reduce_sd(W, U, A, B, C) \ 1023 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1024 (__v2df)(__m128d)(B), \ 1025 (__v2df)(__m128d)(W), (__mmask8)(U), \ 1026 (int)(C), _MM_FROUND_CUR_DIRECTION) 1028 #define _mm_maskz_reduce_sd(U, A, B, C) \ 1029 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1030 (__v2df)(__m128d)(B), \ 1031 (__v2df)_mm_setzero_pd(), \ 1032 (__mmask8)(U), (int)(C), \ 1033 _MM_FROUND_CUR_DIRECTION) 1035 #define _mm_reduce_round_sd(A, B, C, R) \ 1036 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1037 (__v2df)(__m128d)(B), \ 1038 (__v2df)_mm_setzero_pd(), \ 1039 (__mmask8)-1, (int)(C), (int)(R)) 1041 #define _mm_mask_reduce_round_sd(W, U, A, B, C, R) \ 1042 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1043 (__v2df)(__m128d)(B), \ 1044 (__v2df)(__m128d)(W), (__mmask8)(U), \ 1047 #define _mm_maskz_reduce_round_sd(U, A, B, C, R) \ 1048 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1049 (__v2df)(__m128d)(B), \ 1050 (__v2df)_mm_setzero_pd(), \ 1051 (__mmask8)(U), (int)(C), (int)(R)) 1056 return (
__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A);
1062 return (__m512i) __builtin_ia32_cvtmask2d512 (__A);
1068 return (__m512i) __builtin_ia32_cvtmask2q512 (__A);
1074 return (
__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A);
1081 return (__m512)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A,
1082 0, 1, 0, 1, 0, 1, 0, 1,
1083 0, 1, 0, 1, 0, 1, 0, 1);
1089 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1097 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1105 return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A,
1106 0, 1, 2, 3, 4, 5, 6, 7,
1107 0, 1, 2, 3, 4, 5, 6, 7);
1113 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1121 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1129 return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A,
1130 0, 1, 0, 1, 0, 1, 0, 1);
1136 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__M,
1144 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__M,
1152 return (__m512i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
1153 0, 1, 0, 1, 0, 1, 0, 1,
1154 0, 1, 0, 1, 0, 1, 0, 1);
1160 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1168 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1176 return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A,
1177 0, 1, 2, 3, 4, 5, 6, 7,
1178 0, 1, 2, 3, 4, 5, 6, 7);
1184 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1192 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1200 return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A,
1201 0, 1, 0, 1, 0, 1, 0, 1);
1207 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__M,
1215 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__M,
1220 #define _mm512_extractf32x8_ps(A, imm) \ 1221 (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ 1222 (__v8sf)_mm256_undefined_ps(), \ 1225 #define _mm512_mask_extractf32x8_ps(W, U, A, imm) \ 1226 (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ 1227 (__v8sf)(__m256)(W), \ 1230 #define _mm512_maskz_extractf32x8_ps(U, A, imm) \ 1231 (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ 1232 (__v8sf)_mm256_setzero_ps(), \ 1235 #define _mm512_extractf64x2_pd(A, imm) \ 1236 (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ 1238 (__v2df)_mm_undefined_pd(), \ 1241 #define _mm512_mask_extractf64x2_pd(W, U, A, imm) \ 1242 (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ 1244 (__v2df)(__m128d)(W), \ 1247 #define _mm512_maskz_extractf64x2_pd(U, A, imm) \ 1248 (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ 1250 (__v2df)_mm_setzero_pd(), \ 1253 #define _mm512_extracti32x8_epi32(A, imm) \ 1254 (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ 1255 (__v8si)_mm256_undefined_si256(), \ 1258 #define _mm512_mask_extracti32x8_epi32(W, U, A, imm) \ 1259 (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ 1260 (__v8si)(__m256i)(W), \ 1263 #define _mm512_maskz_extracti32x8_epi32(U, A, imm) \ 1264 (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ 1265 (__v8si)_mm256_setzero_si256(), \ 1268 #define _mm512_extracti64x2_epi64(A, imm) \ 1269 (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ 1271 (__v2di)_mm_undefined_si128(), \ 1274 #define _mm512_mask_extracti64x2_epi64(W, U, A, imm) \ 1275 (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ 1277 (__v2di)(__m128i)(W), \ 1280 #define _mm512_maskz_extracti64x2_epi64(U, A, imm) \ 1281 (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ 1283 (__v2di)_mm_setzero_si128(), \ 1286 #define _mm512_insertf32x8(A, B, imm) \ 1287 (__m512)__builtin_ia32_insertf32x8((__v16sf)(__m512)(A), \ 1288 (__v8sf)(__m256)(B), (int)(imm)) 1290 #define _mm512_mask_insertf32x8(W, U, A, B, imm) \ 1291 (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ 1292 (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ 1293 (__v16sf)(__m512)(W)) 1295 #define _mm512_maskz_insertf32x8(U, A, B, imm) \ 1296 (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ 1297 (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ 1298 (__v16sf)_mm512_setzero_ps()) 1300 #define _mm512_insertf64x2(A, B, imm) \ 1301 (__m512d)__builtin_ia32_insertf64x2_512((__v8df)(__m512d)(A), \ 1302 (__v2df)(__m128d)(B), (int)(imm)) 1304 #define _mm512_mask_insertf64x2(W, U, A, B, imm) \ 1305 (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ 1306 (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ 1307 (__v8df)(__m512d)(W)) 1309 #define _mm512_maskz_insertf64x2(U, A, B, imm) \ 1310 (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ 1311 (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ 1312 (__v8df)_mm512_setzero_pd()) 1314 #define _mm512_inserti32x8(A, B, imm) \ 1315 (__m512i)__builtin_ia32_inserti32x8((__v16si)(__m512i)(A), \ 1316 (__v8si)(__m256i)(B), (int)(imm)) 1318 #define _mm512_mask_inserti32x8(W, U, A, B, imm) \ 1319 (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ 1320 (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ 1321 (__v16si)(__m512i)(W)) 1323 #define _mm512_maskz_inserti32x8(U, A, B, imm) \ 1324 (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ 1325 (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ 1326 (__v16si)_mm512_setzero_si512()) 1328 #define _mm512_inserti64x2(A, B, imm) \ 1329 (__m512i)__builtin_ia32_inserti64x2_512((__v8di)(__m512i)(A), \ 1330 (__v2di)(__m128i)(B), (int)(imm)) 1332 #define _mm512_mask_inserti64x2(W, U, A, B, imm) \ 1333 (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ 1334 (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ 1335 (__v8di)(__m512i)(W)) 1337 #define _mm512_maskz_inserti64x2(U, A, B, imm) \ 1338 (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ 1339 (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ 1340 (__v8di)_mm512_setzero_si512()) 1342 #define _mm512_mask_fpclass_ps_mask(U, A, imm) \ 1343 (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ 1344 (int)(imm), (__mmask16)(U)) 1346 #define _mm512_fpclass_ps_mask(A, imm) \ 1347 (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ 1348 (int)(imm), (__mmask16)-1) 1350 #define _mm512_mask_fpclass_pd_mask(U, A, imm) \ 1351 (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ 1354 #define _mm512_fpclass_pd_mask(A, imm) \ 1355 (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ 1358 #define _mm_fpclass_sd_mask(A, imm) \ 1359 (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ 1362 #define _mm_mask_fpclass_sd_mask(U, A, imm) \ 1363 (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ 1366 #define _mm_fpclass_ss_mask(A, imm) \ 1367 (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ 1370 #define _mm_mask_fpclass_ss_mask(U, A, imm) \ 1371 (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ 1374 #undef __DEFAULT_FN_ATTRS512 1375 #undef __DEFAULT_FN_ATTRS static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktestc_mask16_u8(__mmask16 __A, __mmask16 __B)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcast_f32x2(__mmask16 __M, __m128 __A)
static __inline __m256 __DEFAULT_FN_ATTRS _mm256_setzero_ps(void)
Constructs a 256-bit floating-point vector of [8 x float] with all vector elements initialized to zer...
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_broadcast_f64x2(__m128d __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_and_ps(__m512 __A, __m512 __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_andnot_pd(__m512d __A, __m512d __B)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _load_mask8(__mmask8 *__A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_or_ps(__mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_or_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_xor_ps(__m512 __A, __m512 __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_broadcast_f64x2(__m512d __O, __mmask8 __M, __m128d __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcast_f64x2(__mmask8 __M, __m128d __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_and_ps(__mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvttpd_epu64(__mmask8 __U, __m512d __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_xor_ps(__mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_broadcast_i32x8(__m256i __A)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _kadd_mask8(__mmask8 __A, __mmask8 __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepi64_pd(__mmask8 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_broadcast_i64x2(__m128i __A)
static __inline __m512 __DEFAULT_FN_ATTRS512 _mm512_setzero_ps(void)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_and_pd(__m512d __A, __m512d __B)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _kxnor_mask8(__mmask8 __A, __mmask8 __B)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_andnot_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_or_ps(__m512 __A, __m512 __B)
static __inline __m512d __DEFAULT_FN_ATTRS512 _mm512_setzero_pd(void)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvttpd_epu64(__m512d __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_andnot_ps(__mmask16 __U, __m512 __A, __m512 __B)
static __inline__ unsigned int __DEFAULT_FN_ATTRS _cvtmask8_u32(__mmask8 __A)
static __inline__ unsigned char int __C
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_and_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_and_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_or_pd(__mmask8 __U, __m512d __A, __m512d __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_broadcast_i32x2(__m128i __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepu64_pd(__mmask8 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_broadcast_i64x2(__m512i __O, __mmask8 __M, __m128i __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktest_mask16_u8(__mmask16 __A, __mmask16 __B, unsigned char *__C)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtpd_epi64(__mmask8 __U, __m512d __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvttps_epu64(__mmask8 __U, __m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_movm_epi64(__mmask8 __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktestz_mask16_u8(__mmask16 __A, __mmask16 __B)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_broadcast_f32x2(__m512 __O, __mmask16 __M, __m128 __A)
static __inline __m512i __DEFAULT_FN_ATTRS512 _mm512_setzero_si512(void)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcast_f32x8(__mmask16 __M, __m256 __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _kortestz_mask8_u8(__mmask8 __A, __mmask8 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcast_i32x2(__mmask16 __M, __m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvttps_epi64(__m512i __W, __mmask8 __U, __m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtps_epu64(__m512i __W, __mmask8 __U, __m256 __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktestz_mask8_u8(__mmask8 __A, __mmask8 __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _kortest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvttpd_epi64(__m512d __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_cvtepi64_pd(__m512i __A)
static __inline__ __m256 __DEFAULT_FN_ATTRS512 _mm512_cvtepu64_ps(__m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtps_epu64(__mmask8 __U, __m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_broadcast_i32x8(__m512i __O, __mmask16 __M, __m256i __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_xor_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
#define __DEFAULT_FN_ATTRS
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _kandn_mask8(__mmask8 __A, __mmask8 __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_xor_pd(__mmask8 __U, __m512d __A, __m512d __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvttpd_epu64(__m512i __W, __mmask8 __U, __m512d __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mullo_epi64(__m512i __A, __m512i __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvttpd_epi64(__mmask8 __U, __m512d __A)
static __inline__ __mmask16 __DEFAULT_FN_ATTRS _kadd_mask16(__mmask16 __A, __mmask16 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvttps_epi64(__m256 __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_and_pd(__mmask8 __U, __m512d __A, __m512d __B)
static __inline__ __mmask16 __DEFAULT_FN_ATTRS512 _mm512_movepi32_mask(__m512i __A)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _kortestc_mask8_u8(__mmask8 __A, __mmask8 __B)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _cvtu32_mask8(unsigned int __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_xor_ps(__m512 __W, __mmask16 __U, __m512 __A, __m512 __B)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS512 _mm512_movepi64_mask(__m512i __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_andnot_ps(__m512 __A, __m512 __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktest_mask8_u8(__mmask8 __A, __mmask8 __B, unsigned char *__C)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_mask_broadcast_f32x8(__m512 __O, __mmask16 __M, __m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtpd_epu64(__m512i __W, __mmask8 __U, __m512d __A)
static __inline__ __m256 __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepi64_ps(__mmask8 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtps_epi64(__m512i __W, __mmask8 __U, __m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvtpd_epi64(__m512i __W, __mmask8 __U, __m512d __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_mullo_epi64(__m512i __W, __mmask8 __U, __m512i __A, __m512i __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepu64_pd(__m512d __W, __mmask8 __U, __m512i __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_andnot_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _kor_mask8(__mmask8 __A, __mmask8 __B)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _kxor_mask8(__mmask8 __A, __mmask8 __B)
static __inline __mmask8 __DEFAULT_FN_ATTRS _knot_mask8(__mmask8 __M)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvttps_epu64(__m512i __W, __mmask8 __U, __m256 __A)
static __inline__ __mmask8 __DEFAULT_FN_ATTRS _kand_mask8(__mmask8 __A, __mmask8 __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_maskz_andnot_pd(__mmask8 __U, __m512d __A, __m512d __B)
#define _MM_FROUND_CUR_DIRECTION
static __inline__ __m256 __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepi64_ps(__m256 __W, __mmask8 __U, __m512i __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_broadcast_f32x2(__m128 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvtps_epu64(__m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_cvttpd_epi64(__m512i __W, __mmask8 __U, __m512d __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_mullo_epi64(__mmask8 __U, __m512i __A, __m512i __B)
static __inline__ unsigned char __DEFAULT_FN_ATTRS _ktestc_mask8_u8(__mmask8 __A, __mmask8 __B)
#define __DEFAULT_FN_ATTRS512
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcast_i64x2(__mmask8 __M, __m128i __A)
static __inline__ __m256 __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtepu64_ps(__mmask8 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtps_epi64(__mmask8 __U, __m256 __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_or_pd(__m512d __A, __m512d __B)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepi64_pd(__m512d __W, __mmask8 __U, __m512i __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_xor_pd(__m512d __A, __m512d __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvtpd_epi64(__m512d __A)
static __inline__ __m256 __DEFAULT_FN_ATTRS512 _mm512_cvtepi64_ps(__m512i __A)
static __inline__ void __DEFAULT_FN_ATTRS _store_mask8(__mmask8 *__A, __mmask8 __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_broadcast_i32x8(__mmask16 __M, __m256i __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_mask_or_pd(__m512d __W, __mmask8 __U, __m512d __A, __m512d __B)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_mask_broadcast_i32x2(__m512i __O, __mmask16 __M, __m128i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvtps_epi64(__m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvttps_epu64(__m256 __A)
static __inline__ __m512 __DEFAULT_FN_ATTRS512 _mm512_broadcast_f32x8(__m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_movm_epi32(__mmask16 __A)
static __inline__ __m256 __DEFAULT_FN_ATTRS512 _mm512_mask_cvtepu64_ps(__m256 __W, __mmask8 __U, __m512i __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_cvtpd_epu64(__m512d __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvttps_epi64(__mmask8 __U, __m256 __A)
static __inline__ __m512i __DEFAULT_FN_ATTRS512 _mm512_maskz_cvtpd_epu64(__mmask8 __U, __m512d __A)
static __inline__ __m512d __DEFAULT_FN_ATTRS512 _mm512_cvtepu64_pd(__m512i __A)