1 1.1 joerg /*===------------- avx512vbmiintrin.h - VBMI intrinsics ------------------=== 2 1.1 joerg * 3 1.1 joerg * 4 1.1 joerg * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. 5 1.1 joerg * See https://llvm.org/LICENSE.txt for license information. 6 1.1 joerg * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception 7 1.1 joerg * 8 1.1 joerg *===-----------------------------------------------------------------------=== 9 1.1 joerg */ 10 1.1 joerg #ifndef __IMMINTRIN_H 11 1.1 joerg #error "Never use <avx512vbmiintrin.h> directly; include <immintrin.h> instead." 12 1.1 joerg #endif 13 1.1 joerg 14 1.1 joerg #ifndef __VBMIINTRIN_H 15 1.1 joerg #define __VBMIINTRIN_H 16 1.1 joerg 17 1.1 joerg /* Define the default attributes for the functions in this file. */ 18 1.1 joerg #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("avx512vbmi"), __min_vector_width__(512))) 19 1.1 joerg 20 1.1 joerg 21 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 22 1.1 joerg _mm512_permutex2var_epi8(__m512i __A, __m512i __I, __m512i __B) 23 1.1 joerg { 24 1.1 joerg return (__m512i)__builtin_ia32_vpermi2varqi512((__v64qi)__A, (__v64qi)__I, 25 1.1 joerg (__v64qi) __B); 26 1.1 joerg } 27 1.1 joerg 28 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 29 1.1 joerg _mm512_mask_permutex2var_epi8(__m512i __A, __mmask64 __U, __m512i __I, 30 1.1 joerg __m512i __B) 31 1.1 joerg { 32 1.1 joerg return (__m512i)__builtin_ia32_selectb_512(__U, 33 1.1 joerg (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), 34 1.1 joerg (__v64qi)__A); 35 1.1 joerg } 36 1.1 joerg 37 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 38 1.1 joerg _mm512_mask2_permutex2var_epi8(__m512i __A, __m512i __I, __mmask64 __U, 39 1.1 joerg __m512i __B) 40 1.1 joerg { 41 1.1 joerg return (__m512i)__builtin_ia32_selectb_512(__U, 42 1.1 joerg (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), 43 1.1 joerg (__v64qi)__I); 44 1.1 joerg } 45 1.1 joerg 46 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 47 1.1 joerg _mm512_maskz_permutex2var_epi8(__mmask64 __U, __m512i __A, __m512i __I, 48 1.1 joerg __m512i __B) 49 1.1 joerg { 50 1.1 joerg return (__m512i)__builtin_ia32_selectb_512(__U, 51 1.1 joerg (__v64qi)_mm512_permutex2var_epi8(__A, __I, __B), 52 1.1 joerg (__v64qi)_mm512_setzero_si512()); 53 1.1 joerg } 54 1.1 joerg 55 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 56 1.1 joerg _mm512_permutexvar_epi8 (__m512i __A, __m512i __B) 57 1.1 joerg { 58 1.1 joerg return (__m512i)__builtin_ia32_permvarqi512((__v64qi) __B, (__v64qi) __A); 59 1.1 joerg } 60 1.1 joerg 61 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 62 1.1 joerg _mm512_maskz_permutexvar_epi8 (__mmask64 __M, __m512i __A, 63 1.1 joerg __m512i __B) 64 1.1 joerg { 65 1.1 joerg return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, 66 1.1 joerg (__v64qi)_mm512_permutexvar_epi8(__A, __B), 67 1.1 joerg (__v64qi)_mm512_setzero_si512()); 68 1.1 joerg } 69 1.1 joerg 70 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 71 1.1 joerg _mm512_mask_permutexvar_epi8 (__m512i __W, __mmask64 __M, __m512i __A, 72 1.1 joerg __m512i __B) 73 1.1 joerg { 74 1.1 joerg return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, 75 1.1 joerg (__v64qi)_mm512_permutexvar_epi8(__A, __B), 76 1.1 joerg (__v64qi)__W); 77 1.1 joerg } 78 1.1 joerg 79 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 80 1.1 joerg _mm512_multishift_epi64_epi8(__m512i __X, __m512i __Y) 81 1.1 joerg { 82 1.1 joerg return (__m512i)__builtin_ia32_vpmultishiftqb512((__v64qi)__X, (__v64qi) __Y); 83 1.1 joerg } 84 1.1 joerg 85 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 86 1.1 joerg _mm512_mask_multishift_epi64_epi8(__m512i __W, __mmask64 __M, __m512i __X, 87 1.1 joerg __m512i __Y) 88 1.1 joerg { 89 1.1 joerg return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, 90 1.1 joerg (__v64qi)_mm512_multishift_epi64_epi8(__X, __Y), 91 1.1 joerg (__v64qi)__W); 92 1.1 joerg } 93 1.1 joerg 94 1.1 joerg static __inline__ __m512i __DEFAULT_FN_ATTRS 95 1.1 joerg _mm512_maskz_multishift_epi64_epi8(__mmask64 __M, __m512i __X, __m512i __Y) 96 1.1 joerg { 97 1.1 joerg return (__m512i)__builtin_ia32_selectb_512((__mmask64)__M, 98 1.1 joerg (__v64qi)_mm512_multishift_epi64_epi8(__X, __Y), 99 1.1 joerg (__v64qi)_mm512_setzero_si512()); 100 1.1 joerg } 101 1.1 joerg 102 1.1 joerg 103 1.1 joerg #undef __DEFAULT_FN_ATTRS 104 1.1 joerg 105 1.1 joerg #endif 106