25 #error "Never use <avx512dqintrin.h> directly; include <immintrin.h> instead." 28 #ifndef __AVX512DQINTRIN_H 29 #define __AVX512DQINTRIN_H 32 #define __DEFAULT_FN_ATTRS512 __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"), __min_vector_width__(512))) 33 #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512dq"))) 38 return __builtin_ia32_knotqi(__M);
74 return (
unsigned char)__builtin_ia32_kortestcqi(__A, __B);
80 return (
unsigned char)__builtin_ia32_kortestzqi(__A, __B);
85 *__C = (
unsigned char)__builtin_ia32_kortestcqi(__A, __B);
86 return (
unsigned char)__builtin_ia32_kortestzqi(__A, __B);
92 return (
unsigned char)__builtin_ia32_ktestcqi(__A, __B);
98 return (
unsigned char)__builtin_ia32_ktestzqi(__A, __B);
103 *__C = (
unsigned char)__builtin_ia32_ktestcqi(__A, __B);
104 return (
unsigned char)__builtin_ia32_ktestzqi(__A, __B);
110 return (
unsigned char)__builtin_ia32_ktestchi(__A, __B);
116 return (
unsigned char)__builtin_ia32_ktestzhi(__A, __B);
121 *__C = (
unsigned char)__builtin_ia32_ktestchi(__A, __B);
122 return (
unsigned char)__builtin_ia32_ktestzhi(__A, __B);
137 #define _kshiftli_mask8(A, I) \ 138 (__mmask8)__builtin_ia32_kshiftliqi((__mmask8)(A), (unsigned int)(I)) 140 #define _kshiftri_mask8(A, I) \ 141 (__mmask8)__builtin_ia32_kshiftriqi((__mmask8)(A), (unsigned int)(I)) 145 return (
unsigned int)__builtin_ia32_kmovb((
__mmask8)__A);
165 return (__m512i) ((__v8du) __A * (__v8du) __B);
170 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__U,
177 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__U,
184 return (__m512d)((__v8du)__A ^ (__v8du)__B);
189 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
196 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
203 return (__m512)((__v16su)__A ^ (__v16su)__B);
208 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
215 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
222 return (__m512d)((__v8du)__A | (__v8du)__B);
227 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
234 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
241 return (__m512)((__v16su)__A | (__v16su)__B);
246 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
253 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
260 return (__m512d)((__v8du)__A & (__v8du)__B);
265 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
272 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
279 return (__m512)((__v16su)__A & (__v16su)__B);
284 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
291 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
298 return (__m512d)(~(__v8du)__A & (__v8du)__B);
303 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
310 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
317 return (__m512)(~(__v16su)__A & (__v16su)__B);
322 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
329 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__U,
336 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
344 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
352 return (__m512i) __builtin_ia32_cvtpd2qq512_mask ((__v8df) __A,
358 #define _mm512_cvt_roundpd_epi64(A, R) \ 359 (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ 360 (__v8di)_mm512_setzero_si512(), \ 361 (__mmask8)-1, (int)(R)) 363 #define _mm512_mask_cvt_roundpd_epi64(W, U, A, R) \ 364 (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ 365 (__v8di)(__m512i)(W), \ 366 (__mmask8)(U), (int)(R)) 368 #define _mm512_maskz_cvt_roundpd_epi64(U, A, R) \ 369 (__m512i)__builtin_ia32_cvtpd2qq512_mask((__v8df)(__m512d)(A), \ 370 (__v8di)_mm512_setzero_si512(), \ 371 (__mmask8)(U), (int)(R)) 375 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
383 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
391 return (__m512i) __builtin_ia32_cvtpd2uqq512_mask ((__v8df) __A,
397 #define _mm512_cvt_roundpd_epu64(A, R) \ 398 (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ 399 (__v8di)_mm512_setzero_si512(), \ 400 (__mmask8)-1, (int)(R)) 402 #define _mm512_mask_cvt_roundpd_epu64(W, U, A, R) \ 403 (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ 404 (__v8di)(__m512i)(W), \ 405 (__mmask8)(U), (int)(R)) 407 #define _mm512_maskz_cvt_roundpd_epu64(U, A, R) \ 408 (__m512i)__builtin_ia32_cvtpd2uqq512_mask((__v8df)(__m512d)(A), \ 409 (__v8di)_mm512_setzero_si512(), \ 410 (__mmask8)(U), (int)(R)) 414 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
422 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
430 return (__m512i) __builtin_ia32_cvtps2qq512_mask ((__v8sf) __A,
436 #define _mm512_cvt_roundps_epi64(A, R) \ 437 (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ 438 (__v8di)_mm512_setzero_si512(), \ 439 (__mmask8)-1, (int)(R)) 441 #define _mm512_mask_cvt_roundps_epi64(W, U, A, R) \ 442 (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ 443 (__v8di)(__m512i)(W), \ 444 (__mmask8)(U), (int)(R)) 446 #define _mm512_maskz_cvt_roundps_epi64(U, A, R) \ 447 (__m512i)__builtin_ia32_cvtps2qq512_mask((__v8sf)(__m256)(A), \ 448 (__v8di)_mm512_setzero_si512(), \ 449 (__mmask8)(U), (int)(R)) 453 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
461 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
469 return (__m512i) __builtin_ia32_cvtps2uqq512_mask ((__v8sf) __A,
475 #define _mm512_cvt_roundps_epu64(A, R) \ 476 (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ 477 (__v8di)_mm512_setzero_si512(), \ 478 (__mmask8)-1, (int)(R)) 480 #define _mm512_mask_cvt_roundps_epu64(W, U, A, R) \ 481 (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ 482 (__v8di)(__m512i)(W), \ 483 (__mmask8)(U), (int)(R)) 485 #define _mm512_maskz_cvt_roundps_epu64(U, A, R) \ 486 (__m512i)__builtin_ia32_cvtps2uqq512_mask((__v8sf)(__m256)(A), \ 487 (__v8di)_mm512_setzero_si512(), \ 488 (__mmask8)(U), (int)(R)) 493 return (__m512d)__builtin_convertvector((__v8di)__A, __v8df);
498 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
505 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
510 #define _mm512_cvt_roundepi64_pd(A, R) \ 511 (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ 512 (__v8df)_mm512_setzero_pd(), \ 513 (__mmask8)-1, (int)(R)) 515 #define _mm512_mask_cvt_roundepi64_pd(W, U, A, R) \ 516 (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ 517 (__v8df)(__m512d)(W), \ 518 (__mmask8)(U), (int)(R)) 520 #define _mm512_maskz_cvt_roundepi64_pd(U, A, R) \ 521 (__m512d)__builtin_ia32_cvtqq2pd512_mask((__v8di)(__m512i)(A), \ 522 (__v8df)_mm512_setzero_pd(), \ 523 (__mmask8)(U), (int)(R)) 527 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
535 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
543 return (__m256) __builtin_ia32_cvtqq2ps512_mask ((__v8di) __A,
549 #define _mm512_cvt_roundepi64_ps(A, R) \ 550 (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ 551 (__v8sf)_mm256_setzero_ps(), \ 552 (__mmask8)-1, (int)(R)) 554 #define _mm512_mask_cvt_roundepi64_ps(W, U, A, R) \ 555 (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ 556 (__v8sf)(__m256)(W), (__mmask8)(U), \ 559 #define _mm512_maskz_cvt_roundepi64_ps(U, A, R) \ 560 (__m256)__builtin_ia32_cvtqq2ps512_mask((__v8di)(__m512i)(A), \ 561 (__v8sf)_mm256_setzero_ps(), \ 562 (__mmask8)(U), (int)(R)) 567 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
575 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
583 return (__m512i) __builtin_ia32_cvttpd2qq512_mask ((__v8df) __A,
589 #define _mm512_cvtt_roundpd_epi64(A, R) \ 590 (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ 591 (__v8di)_mm512_setzero_si512(), \ 592 (__mmask8)-1, (int)(R)) 594 #define _mm512_mask_cvtt_roundpd_epi64(W, U, A, R) \ 595 (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ 596 (__v8di)(__m512i)(W), \ 597 (__mmask8)(U), (int)(R)) 599 #define _mm512_maskz_cvtt_roundpd_epi64(U, A, R) \ 600 (__m512i)__builtin_ia32_cvttpd2qq512_mask((__v8df)(__m512d)(A), \ 601 (__v8di)_mm512_setzero_si512(), \ 602 (__mmask8)(U), (int)(R)) 606 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
614 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
622 return (__m512i) __builtin_ia32_cvttpd2uqq512_mask ((__v8df) __A,
628 #define _mm512_cvtt_roundpd_epu64(A, R) \ 629 (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ 630 (__v8di)_mm512_setzero_si512(), \ 631 (__mmask8)-1, (int)(R)) 633 #define _mm512_mask_cvtt_roundpd_epu64(W, U, A, R) \ 634 (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ 635 (__v8di)(__m512i)(W), \ 636 (__mmask8)(U), (int)(R)) 638 #define _mm512_maskz_cvtt_roundpd_epu64(U, A, R) \ 639 (__m512i)__builtin_ia32_cvttpd2uqq512_mask((__v8df)(__m512d)(A), \ 640 (__v8di)_mm512_setzero_si512(), \ 641 (__mmask8)(U), (int)(R)) 645 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
653 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
661 return (__m512i) __builtin_ia32_cvttps2qq512_mask ((__v8sf) __A,
667 #define _mm512_cvtt_roundps_epi64(A, R) \ 668 (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ 669 (__v8di)_mm512_setzero_si512(), \ 670 (__mmask8)-1, (int)(R)) 672 #define _mm512_mask_cvtt_roundps_epi64(W, U, A, R) \ 673 (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ 674 (__v8di)(__m512i)(W), \ 675 (__mmask8)(U), (int)(R)) 677 #define _mm512_maskz_cvtt_roundps_epi64(U, A, R) \ 678 (__m512i)__builtin_ia32_cvttps2qq512_mask((__v8sf)(__m256)(A), \ 679 (__v8di)_mm512_setzero_si512(), \ 680 (__mmask8)(U), (int)(R)) 684 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
692 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
700 return (__m512i) __builtin_ia32_cvttps2uqq512_mask ((__v8sf) __A,
706 #define _mm512_cvtt_roundps_epu64(A, R) \ 707 (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ 708 (__v8di)_mm512_setzero_si512(), \ 709 (__mmask8)-1, (int)(R)) 711 #define _mm512_mask_cvtt_roundps_epu64(W, U, A, R) \ 712 (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ 713 (__v8di)(__m512i)(W), \ 714 (__mmask8)(U), (int)(R)) 716 #define _mm512_maskz_cvtt_roundps_epu64(U, A, R) \ 717 (__m512i)__builtin_ia32_cvttps2uqq512_mask((__v8sf)(__m256)(A), \ 718 (__v8di)_mm512_setzero_si512(), \ 719 (__mmask8)(U), (int)(R)) 723 return (__m512d)__builtin_convertvector((__v8du)__A, __v8df);
728 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
735 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__U,
740 #define _mm512_cvt_roundepu64_pd(A, R) \ 741 (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ 742 (__v8df)_mm512_setzero_pd(), \ 743 (__mmask8)-1, (int)(R)) 745 #define _mm512_mask_cvt_roundepu64_pd(W, U, A, R) \ 746 (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ 747 (__v8df)(__m512d)(W), \ 748 (__mmask8)(U), (int)(R)) 751 #define _mm512_maskz_cvt_roundepu64_pd(U, A, R) \ 752 (__m512d)__builtin_ia32_cvtuqq2pd512_mask((__v8di)(__m512i)(A), \ 753 (__v8df)_mm512_setzero_pd(), \ 754 (__mmask8)(U), (int)(R)) 759 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
767 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
775 return (__m256) __builtin_ia32_cvtuqq2ps512_mask ((__v8di) __A,
781 #define _mm512_cvt_roundepu64_ps(A, R) \ 782 (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ 783 (__v8sf)_mm256_setzero_ps(), \ 784 (__mmask8)-1, (int)(R)) 786 #define _mm512_mask_cvt_roundepu64_ps(W, U, A, R) \ 787 (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ 788 (__v8sf)(__m256)(W), (__mmask8)(U), \ 791 #define _mm512_maskz_cvt_roundepu64_ps(U, A, R) \ 792 (__m256)__builtin_ia32_cvtuqq2ps512_mask((__v8di)(__m512i)(A), \ 793 (__v8sf)_mm256_setzero_ps(), \ 794 (__mmask8)(U), (int)(R)) 796 #define _mm512_range_pd(A, B, C) \ 797 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 798 (__v8df)(__m512d)(B), (int)(C), \ 799 (__v8df)_mm512_setzero_pd(), \ 801 _MM_FROUND_CUR_DIRECTION) 803 #define _mm512_mask_range_pd(W, U, A, B, C) \ 804 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 805 (__v8df)(__m512d)(B), (int)(C), \ 806 (__v8df)(__m512d)(W), (__mmask8)(U), \ 807 _MM_FROUND_CUR_DIRECTION) 809 #define _mm512_maskz_range_pd(U, A, B, C) \ 810 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 811 (__v8df)(__m512d)(B), (int)(C), \ 812 (__v8df)_mm512_setzero_pd(), \ 814 _MM_FROUND_CUR_DIRECTION) 816 #define _mm512_range_round_pd(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)-1, (int)(R)) 822 #define _mm512_mask_range_round_pd(W, U, A, B, C, R) \ 823 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 824 (__v8df)(__m512d)(B), (int)(C), \ 825 (__v8df)(__m512d)(W), (__mmask8)(U), \ 828 #define _mm512_maskz_range_round_pd(U, A, B, C, R) \ 829 (__m512d)__builtin_ia32_rangepd512_mask((__v8df)(__m512d)(A), \ 830 (__v8df)(__m512d)(B), (int)(C), \ 831 (__v8df)_mm512_setzero_pd(), \ 832 (__mmask8)(U), (int)(R)) 834 #define _mm512_range_ps(A, B, C) \ 835 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 836 (__v16sf)(__m512)(B), (int)(C), \ 837 (__v16sf)_mm512_setzero_ps(), \ 839 _MM_FROUND_CUR_DIRECTION) 841 #define _mm512_mask_range_ps(W, U, A, B, C) \ 842 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 843 (__v16sf)(__m512)(B), (int)(C), \ 844 (__v16sf)(__m512)(W), (__mmask16)(U), \ 845 _MM_FROUND_CUR_DIRECTION) 847 #define _mm512_maskz_range_ps(U, A, B, C) \ 848 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 849 (__v16sf)(__m512)(B), (int)(C), \ 850 (__v16sf)_mm512_setzero_ps(), \ 852 _MM_FROUND_CUR_DIRECTION) 854 #define _mm512_range_round_ps(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)-1, (int)(R)) 860 #define _mm512_mask_range_round_ps(W, U, A, B, C, R) \ 861 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 862 (__v16sf)(__m512)(B), (int)(C), \ 863 (__v16sf)(__m512)(W), (__mmask16)(U), \ 866 #define _mm512_maskz_range_round_ps(U, A, B, C, R) \ 867 (__m512)__builtin_ia32_rangeps512_mask((__v16sf)(__m512)(A), \ 868 (__v16sf)(__m512)(B), (int)(C), \ 869 (__v16sf)_mm512_setzero_ps(), \ 870 (__mmask16)(U), (int)(R)) 872 #define _mm_range_round_ss(A, B, C, R) \ 873 (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ 874 (__v4sf)(__m128)(B), \ 875 (__v4sf)_mm_setzero_ps(), \ 876 (__mmask8) -1, (int)(C),\ 879 #define _mm_range_ss(A ,B , C) _mm_range_round_ss(A, B, C ,_MM_FROUND_CUR_DIRECTION) 881 #define _mm_mask_range_round_ss(W, U, A, B, C, R) \ 882 (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ 883 (__v4sf)(__m128)(B), \ 884 (__v4sf)(__m128)(W),\ 885 (__mmask8)(U), (int)(C),\ 888 #define _mm_mask_range_ss(W , U, A, B, C) _mm_mask_range_round_ss(W, U, A, B, C , _MM_FROUND_CUR_DIRECTION) 890 #define _mm_maskz_range_round_ss(U, A, B, C, R) \ 891 (__m128)__builtin_ia32_rangess128_round_mask((__v4sf)(__m128)(A), \ 892 (__v4sf)(__m128)(B), \ 893 (__v4sf)_mm_setzero_ps(), \ 894 (__mmask8)(U), (int)(C),\ 897 #define _mm_maskz_range_ss(U, A ,B , C) _mm_maskz_range_round_ss(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) 899 #define _mm_range_round_sd(A, B, C, R) \ 900 (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ 901 (__v2df)(__m128d)(B), \ 902 (__v2df)_mm_setzero_pd(), \ 903 (__mmask8) -1, (int)(C),\ 906 #define _mm_range_sd(A ,B , C) _mm_range_round_sd(A, B, C ,_MM_FROUND_CUR_DIRECTION) 908 #define _mm_mask_range_round_sd(W, U, A, B, C, R) \ 909 (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ 910 (__v2df)(__m128d)(B), \ 911 (__v2df)(__m128d)(W),\ 912 (__mmask8)(U), (int)(C),\ 915 #define _mm_mask_range_sd(W, U, A, B, C) _mm_mask_range_round_sd(W, U, A, B, C ,_MM_FROUND_CUR_DIRECTION) 917 #define _mm_maskz_range_round_sd(U, A, B, C, R) \ 918 (__m128d)__builtin_ia32_rangesd128_round_mask((__v2df)(__m128d)(A), \ 919 (__v2df)(__m128d)(B), \ 920 (__v2df)_mm_setzero_pd(), \ 921 (__mmask8)(U), (int)(C),\ 924 #define _mm_maskz_range_sd(U, A, B, C) _mm_maskz_range_round_sd(U, A, B, C ,_MM_FROUND_CUR_DIRECTION) 926 #define _mm512_reduce_pd(A, B) \ 927 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 928 (__v8df)_mm512_setzero_pd(), \ 930 _MM_FROUND_CUR_DIRECTION) 932 #define _mm512_mask_reduce_pd(W, U, A, B) \ 933 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 934 (__v8df)(__m512d)(W), \ 936 _MM_FROUND_CUR_DIRECTION) 938 #define _mm512_maskz_reduce_pd(U, A, B) \ 939 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 940 (__v8df)_mm512_setzero_pd(), \ 942 _MM_FROUND_CUR_DIRECTION) 944 #define _mm512_reduce_ps(A, B) \ 945 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 946 (__v16sf)_mm512_setzero_ps(), \ 948 _MM_FROUND_CUR_DIRECTION) 950 #define _mm512_mask_reduce_ps(W, U, A, B) \ 951 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 952 (__v16sf)(__m512)(W), \ 954 _MM_FROUND_CUR_DIRECTION) 956 #define _mm512_maskz_reduce_ps(U, A, B) \ 957 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 958 (__v16sf)_mm512_setzero_ps(), \ 960 _MM_FROUND_CUR_DIRECTION) 962 #define _mm512_reduce_round_pd(A, B, R) \ 963 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 964 (__v8df)_mm512_setzero_pd(), \ 965 (__mmask8)-1, (int)(R)) 967 #define _mm512_mask_reduce_round_pd(W, U, A, B, R) \ 968 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 969 (__v8df)(__m512d)(W), \ 970 (__mmask8)(U), (int)(R)) 972 #define _mm512_maskz_reduce_round_pd(U, A, B, R) \ 973 (__m512d)__builtin_ia32_reducepd512_mask((__v8df)(__m512d)(A), (int)(B), \ 974 (__v8df)_mm512_setzero_pd(), \ 975 (__mmask8)(U), (int)(R)) 977 #define _mm512_reduce_round_ps(A, B, R) \ 978 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 979 (__v16sf)_mm512_setzero_ps(), \ 980 (__mmask16)-1, (int)(R)) 982 #define _mm512_mask_reduce_round_ps(W, U, A, B, R) \ 983 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 984 (__v16sf)(__m512)(W), \ 985 (__mmask16)(U), (int)(R)) 987 #define _mm512_maskz_reduce_round_ps(U, A, B, R) \ 988 (__m512)__builtin_ia32_reduceps512_mask((__v16sf)(__m512)(A), (int)(B), \ 989 (__v16sf)_mm512_setzero_ps(), \ 990 (__mmask16)(U), (int)(R)) 992 #define _mm_reduce_ss(A, B, C) \ 993 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 994 (__v4sf)(__m128)(B), \ 995 (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ 996 (int)(C), _MM_FROUND_CUR_DIRECTION) 998 #define _mm_mask_reduce_ss(W, U, A, B, C) \ 999 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1000 (__v4sf)(__m128)(B), \ 1001 (__v4sf)(__m128)(W), (__mmask8)(U), \ 1002 (int)(C), _MM_FROUND_CUR_DIRECTION) 1004 #define _mm_maskz_reduce_ss(U, A, B, C) \ 1005 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1006 (__v4sf)(__m128)(B), \ 1007 (__v4sf)_mm_setzero_ps(), \ 1008 (__mmask8)(U), (int)(C), \ 1009 _MM_FROUND_CUR_DIRECTION) 1011 #define _mm_reduce_round_ss(A, B, C, R) \ 1012 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1013 (__v4sf)(__m128)(B), \ 1014 (__v4sf)_mm_setzero_ps(), (__mmask8)-1, \ 1017 #define _mm_mask_reduce_round_ss(W, U, A, B, C, R) \ 1018 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1019 (__v4sf)(__m128)(B), \ 1020 (__v4sf)(__m128)(W), (__mmask8)(U), \ 1023 #define _mm_maskz_reduce_round_ss(U, A, B, C, R) \ 1024 (__m128)__builtin_ia32_reducess_mask((__v4sf)(__m128)(A), \ 1025 (__v4sf)(__m128)(B), \ 1026 (__v4sf)_mm_setzero_ps(), \ 1027 (__mmask8)(U), (int)(C), (int)(R)) 1029 #define _mm_reduce_sd(A, B, C) \ 1030 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1031 (__v2df)(__m128d)(B), \ 1032 (__v2df)_mm_setzero_pd(), \ 1033 (__mmask8)-1, (int)(C), \ 1034 _MM_FROUND_CUR_DIRECTION) 1036 #define _mm_mask_reduce_sd(W, U, A, B, C) \ 1037 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1038 (__v2df)(__m128d)(B), \ 1039 (__v2df)(__m128d)(W), (__mmask8)(U), \ 1040 (int)(C), _MM_FROUND_CUR_DIRECTION) 1042 #define _mm_maskz_reduce_sd(U, A, B, C) \ 1043 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1044 (__v2df)(__m128d)(B), \ 1045 (__v2df)_mm_setzero_pd(), \ 1046 (__mmask8)(U), (int)(C), \ 1047 _MM_FROUND_CUR_DIRECTION) 1049 #define _mm_reduce_round_sd(A, B, C, R) \ 1050 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1051 (__v2df)(__m128d)(B), \ 1052 (__v2df)_mm_setzero_pd(), \ 1053 (__mmask8)-1, (int)(C), (int)(R)) 1055 #define _mm_mask_reduce_round_sd(W, U, A, B, C, R) \ 1056 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1057 (__v2df)(__m128d)(B), \ 1058 (__v2df)(__m128d)(W), (__mmask8)(U), \ 1061 #define _mm_maskz_reduce_round_sd(U, A, B, C, R) \ 1062 (__m128d)__builtin_ia32_reducesd_mask((__v2df)(__m128d)(A), \ 1063 (__v2df)(__m128d)(B), \ 1064 (__v2df)_mm_setzero_pd(), \ 1065 (__mmask8)(U), (int)(C), (int)(R)) 1070 return (
__mmask16) __builtin_ia32_cvtd2mask512 ((__v16si) __A);
1076 return (__m512i) __builtin_ia32_cvtmask2d512 (__A);
1082 return (__m512i) __builtin_ia32_cvtmask2q512 (__A);
1088 return (
__mmask8) __builtin_ia32_cvtq2mask512 ((__v8di) __A);
1095 return (__m512)__builtin_shufflevector((__v4sf)__A, (__v4sf)__A,
1096 0, 1, 0, 1, 0, 1, 0, 1,
1097 0, 1, 0, 1, 0, 1, 0, 1);
1103 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1111 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1119 return (__m512)__builtin_shufflevector((__v8sf)__A, (__v8sf)__A,
1120 0, 1, 2, 3, 4, 5, 6, 7,
1121 0, 1, 2, 3, 4, 5, 6, 7);
1127 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1135 return (__m512)__builtin_ia32_selectps_512((
__mmask16)__M,
1143 return (__m512d)__builtin_shufflevector((__v2df)__A, (__v2df)__A,
1144 0, 1, 0, 1, 0, 1, 0, 1);
1150 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__M,
1158 return (__m512d)__builtin_ia32_selectpd_512((
__mmask8)__M,
1166 return (__m512i)__builtin_shufflevector((__v4si)__A, (__v4si)__A,
1167 0, 1, 0, 1, 0, 1, 0, 1,
1168 0, 1, 0, 1, 0, 1, 0, 1);
1174 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1182 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1190 return (__m512i)__builtin_shufflevector((__v8si)__A, (__v8si)__A,
1191 0, 1, 2, 3, 4, 5, 6, 7,
1192 0, 1, 2, 3, 4, 5, 6, 7);
1198 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1206 return (__m512i)__builtin_ia32_selectd_512((
__mmask16)__M,
1214 return (__m512i)__builtin_shufflevector((__v2di)__A, (__v2di)__A,
1215 0, 1, 0, 1, 0, 1, 0, 1);
1221 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__M,
1229 return (__m512i)__builtin_ia32_selectq_512((
__mmask8)__M,
1234 #define _mm512_extractf32x8_ps(A, imm) \ 1235 (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ 1236 (__v8sf)_mm256_undefined_ps(), \ 1239 #define _mm512_mask_extractf32x8_ps(W, U, A, imm) \ 1240 (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ 1241 (__v8sf)(__m256)(W), \ 1244 #define _mm512_maskz_extractf32x8_ps(U, A, imm) \ 1245 (__m256)__builtin_ia32_extractf32x8_mask((__v16sf)(__m512)(A), (int)(imm), \ 1246 (__v8sf)_mm256_setzero_ps(), \ 1249 #define _mm512_extractf64x2_pd(A, imm) \ 1250 (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ 1252 (__v2df)_mm_undefined_pd(), \ 1255 #define _mm512_mask_extractf64x2_pd(W, U, A, imm) \ 1256 (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ 1258 (__v2df)(__m128d)(W), \ 1261 #define _mm512_maskz_extractf64x2_pd(U, A, imm) \ 1262 (__m128d)__builtin_ia32_extractf64x2_512_mask((__v8df)(__m512d)(A), \ 1264 (__v2df)_mm_setzero_pd(), \ 1267 #define _mm512_extracti32x8_epi32(A, imm) \ 1268 (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ 1269 (__v8si)_mm256_undefined_si256(), \ 1272 #define _mm512_mask_extracti32x8_epi32(W, U, A, imm) \ 1273 (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ 1274 (__v8si)(__m256i)(W), \ 1277 #define _mm512_maskz_extracti32x8_epi32(U, A, imm) \ 1278 (__m256i)__builtin_ia32_extracti32x8_mask((__v16si)(__m512i)(A), (int)(imm), \ 1279 (__v8si)_mm256_setzero_si256(), \ 1282 #define _mm512_extracti64x2_epi64(A, imm) \ 1283 (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ 1285 (__v2di)_mm_undefined_si128(), \ 1288 #define _mm512_mask_extracti64x2_epi64(W, U, A, imm) \ 1289 (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ 1291 (__v2di)(__m128i)(W), \ 1294 #define _mm512_maskz_extracti64x2_epi64(U, A, imm) \ 1295 (__m128i)__builtin_ia32_extracti64x2_512_mask((__v8di)(__m512i)(A), \ 1297 (__v2di)_mm_setzero_si128(), \ 1300 #define _mm512_insertf32x8(A, B, imm) \ 1301 (__m512)__builtin_ia32_insertf32x8((__v16sf)(__m512)(A), \ 1302 (__v8sf)(__m256)(B), (int)(imm)) 1304 #define _mm512_mask_insertf32x8(W, U, A, B, imm) \ 1305 (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ 1306 (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ 1307 (__v16sf)(__m512)(W)) 1309 #define _mm512_maskz_insertf32x8(U, A, B, imm) \ 1310 (__m512)__builtin_ia32_selectps_512((__mmask16)(U), \ 1311 (__v16sf)_mm512_insertf32x8((A), (B), (imm)), \ 1312 (__v16sf)_mm512_setzero_ps()) 1314 #define _mm512_insertf64x2(A, B, imm) \ 1315 (__m512d)__builtin_ia32_insertf64x2_512((__v8df)(__m512d)(A), \ 1316 (__v2df)(__m128d)(B), (int)(imm)) 1318 #define _mm512_mask_insertf64x2(W, U, A, B, imm) \ 1319 (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ 1320 (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ 1321 (__v8df)(__m512d)(W)) 1323 #define _mm512_maskz_insertf64x2(U, A, B, imm) \ 1324 (__m512d)__builtin_ia32_selectpd_512((__mmask8)(U), \ 1325 (__v8df)_mm512_insertf64x2((A), (B), (imm)), \ 1326 (__v8df)_mm512_setzero_pd()) 1328 #define _mm512_inserti32x8(A, B, imm) \ 1329 (__m512i)__builtin_ia32_inserti32x8((__v16si)(__m512i)(A), \ 1330 (__v8si)(__m256i)(B), (int)(imm)) 1332 #define _mm512_mask_inserti32x8(W, U, A, B, imm) \ 1333 (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ 1334 (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ 1335 (__v16si)(__m512i)(W)) 1337 #define _mm512_maskz_inserti32x8(U, A, B, imm) \ 1338 (__m512i)__builtin_ia32_selectd_512((__mmask16)(U), \ 1339 (__v16si)_mm512_inserti32x8((A), (B), (imm)), \ 1340 (__v16si)_mm512_setzero_si512()) 1342 #define _mm512_inserti64x2(A, B, imm) \ 1343 (__m512i)__builtin_ia32_inserti64x2_512((__v8di)(__m512i)(A), \ 1344 (__v2di)(__m128i)(B), (int)(imm)) 1346 #define _mm512_mask_inserti64x2(W, U, A, B, imm) \ 1347 (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ 1348 (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ 1349 (__v8di)(__m512i)(W)) 1351 #define _mm512_maskz_inserti64x2(U, A, B, imm) \ 1352 (__m512i)__builtin_ia32_selectq_512((__mmask8)(U), \ 1353 (__v8di)_mm512_inserti64x2((A), (B), (imm)), \ 1354 (__v8di)_mm512_setzero_si512()) 1356 #define _mm512_mask_fpclass_ps_mask(U, A, imm) \ 1357 (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ 1358 (int)(imm), (__mmask16)(U)) 1360 #define _mm512_fpclass_ps_mask(A, imm) \ 1361 (__mmask16)__builtin_ia32_fpclassps512_mask((__v16sf)(__m512)(A), \ 1362 (int)(imm), (__mmask16)-1) 1364 #define _mm512_mask_fpclass_pd_mask(U, A, imm) \ 1365 (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ 1368 #define _mm512_fpclass_pd_mask(A, imm) \ 1369 (__mmask8)__builtin_ia32_fpclasspd512_mask((__v8df)(__m512d)(A), (int)(imm), \ 1372 #define _mm_fpclass_sd_mask(A, imm) \ 1373 (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ 1376 #define _mm_mask_fpclass_sd_mask(U, A, imm) \ 1377 (__mmask8)__builtin_ia32_fpclasssd_mask((__v2df)(__m128d)(A), (int)(imm), \ 1380 #define _mm_fpclass_ss_mask(A, imm) \ 1381 (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ 1384 #define _mm_mask_fpclass_ss_mask(U, A, imm) \ 1385 (__mmask8)__builtin_ia32_fpclassss_mask((__v4sf)(__m128)(A), (int)(imm), \ 1388 #undef __DEFAULT_FN_ATTRS512 1389 #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__ __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)