| 1 | /*===------------- avx512vbmivlintrin.h - VBMI intrinsics ------------------=== |
| 2 | * |
| 3 | * |
| 4 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 5 | * See https://llvm.org/LICENSE.txt for license information. |
| 6 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 7 | * |
| 8 | *===-----------------------------------------------------------------------=== |
| 9 | */ |
| 10 | #ifndef __IMMINTRIN_H |
| 11 | #error "Never use <avx512vbmivlintrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif |
| 13 | |
| 14 | #ifndef __VBMIVLINTRIN_H |
| 15 | #define __VBMIVLINTRIN_H |
| 16 | |
| 17 | /* Define the default attributes for the functions in this file. */ |
| 18 | #if defined(__cplusplus) && (__cplusplus >= 201103L) |
| 19 | #define __DEFAULT_FN_ATTRS128 \ |
| 20 | __attribute__((__always_inline__, __nodebug__, \ |
| 21 | __target__("avx512vbmi,avx512vl"), \ |
| 22 | __min_vector_width__(128))) constexpr |
| 23 | #define __DEFAULT_FN_ATTRS256 \ |
| 24 | __attribute__((__always_inline__, __nodebug__, \ |
| 25 | __target__("avx512vbmi,avx512vl"), \ |
| 26 | __min_vector_width__(256))) constexpr |
| 27 | #else |
| 28 | #define __DEFAULT_FN_ATTRS128 \ |
| 29 | __attribute__((__always_inline__, __nodebug__, \ |
| 30 | __target__("avx512vbmi,avx512vl"), \ |
| 31 | __min_vector_width__(128))) |
| 32 | #define __DEFAULT_FN_ATTRS256 \ |
| 33 | __attribute__((__always_inline__, __nodebug__, \ |
| 34 | __target__("avx512vbmi,avx512vl"), \ |
| 35 | __min_vector_width__(256))) |
| 36 | #endif |
| 37 | |
| 38 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 39 | _mm_permutex2var_epi8(__m128i __A, __m128i __I, __m128i __B) { |
| 40 | return (__m128i)__builtin_ia32_vpermi2varqi128((__v16qi)__A, |
| 41 | (__v16qi)__I, |
| 42 | (__v16qi)__B); |
| 43 | } |
| 44 | |
| 45 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutex2var_epi8( |
| 46 | __m128i __A, __mmask16 __U, __m128i __I, __m128i __B) { |
| 47 | return (__m128i)__builtin_ia32_selectb_128(__U, |
| 48 | (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), |
| 49 | (__v16qi)__A); |
| 50 | } |
| 51 | |
| 52 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask2_permutex2var_epi8( |
| 53 | __m128i __A, __m128i __I, __mmask16 __U, __m128i __B) { |
| 54 | return (__m128i)__builtin_ia32_selectb_128(__U, |
| 55 | (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), |
| 56 | (__v16qi)__I); |
| 57 | } |
| 58 | |
| 59 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_maskz_permutex2var_epi8( |
| 60 | __mmask16 __U, __m128i __A, __m128i __I, __m128i __B) { |
| 61 | return (__m128i)__builtin_ia32_selectb_128(__U, |
| 62 | (__v16qi)_mm_permutex2var_epi8(__A, __I, __B), |
| 63 | (__v16qi)_mm_setzero_si128()); |
| 64 | } |
| 65 | |
| 66 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 67 | _mm256_permutex2var_epi8(__m256i __A, __m256i __I, __m256i __B) { |
| 68 | return (__m256i)__builtin_ia32_vpermi2varqi256((__v32qi)__A, (__v32qi)__I, |
| 69 | (__v32qi)__B); |
| 70 | } |
| 71 | |
| 72 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutex2var_epi8( |
| 73 | __m256i __A, __mmask32 __U, __m256i __I, __m256i __B) { |
| 74 | return (__m256i)__builtin_ia32_selectb_256(__U, |
| 75 | (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), |
| 76 | (__v32qi)__A); |
| 77 | } |
| 78 | |
| 79 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask2_permutex2var_epi8( |
| 80 | __m256i __A, __m256i __I, __mmask32 __U, __m256i __B) { |
| 81 | return (__m256i)__builtin_ia32_selectb_256(__U, |
| 82 | (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), |
| 83 | (__v32qi)__I); |
| 84 | } |
| 85 | |
| 86 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_maskz_permutex2var_epi8( |
| 87 | __mmask32 __U, __m256i __A, __m256i __I, __m256i __B) { |
| 88 | return (__m256i)__builtin_ia32_selectb_256(__U, |
| 89 | (__v32qi)_mm256_permutex2var_epi8(__A, __I, __B), |
| 90 | (__v32qi)_mm256_setzero_si256()); |
| 91 | } |
| 92 | |
| 93 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 94 | _mm_permutexvar_epi8(__m128i __A, __m128i __B) { |
| 95 | return (__m128i)__builtin_ia32_permvarqi128((__v16qi)__B, (__v16qi)__A); |
| 96 | } |
| 97 | |
| 98 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 99 | _mm_maskz_permutexvar_epi8(__mmask16 __M, __m128i __A, __m128i __B) { |
| 100 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
| 101 | (__v16qi)_mm_permutexvar_epi8(__A, __B), |
| 102 | (__v16qi)_mm_setzero_si128()); |
| 103 | } |
| 104 | |
| 105 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_mask_permutexvar_epi8( |
| 106 | __m128i __W, __mmask16 __M, __m128i __A, __m128i __B) { |
| 107 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
| 108 | (__v16qi)_mm_permutexvar_epi8(__A, __B), |
| 109 | (__v16qi)__W); |
| 110 | } |
| 111 | |
| 112 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 113 | _mm256_permutexvar_epi8(__m256i __A, __m256i __B) { |
| 114 | return (__m256i)__builtin_ia32_permvarqi256((__v32qi) __B, (__v32qi) __A); |
| 115 | } |
| 116 | |
| 117 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 118 | _mm256_maskz_permutexvar_epi8(__mmask32 __M, __m256i __A, __m256i __B) { |
| 119 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
| 120 | (__v32qi)_mm256_permutexvar_epi8(__A, __B), |
| 121 | (__v32qi)_mm256_setzero_si256()); |
| 122 | } |
| 123 | |
| 124 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 _mm256_mask_permutexvar_epi8( |
| 125 | __m256i __W, __mmask32 __M, __m256i __A, __m256i __B) { |
| 126 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
| 127 | (__v32qi)_mm256_permutexvar_epi8(__A, __B), |
| 128 | (__v32qi)__W); |
| 129 | } |
| 130 | |
| 131 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 132 | _mm_multishift_epi64_epi8(__m128i __X, __m128i __Y) |
| 133 | { |
| 134 | return (__m128i)__builtin_ia32_vpmultishiftqb128((__v16qi)__X, (__v16qi)__Y); |
| 135 | } |
| 136 | |
| 137 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 138 | _mm_mask_multishift_epi64_epi8(__m128i __W, __mmask16 __M, __m128i __X, |
| 139 | __m128i __Y) |
| 140 | { |
| 141 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
| 142 | (__v16qi)_mm_multishift_epi64_epi8(__X, __Y), |
| 143 | (__v16qi)__W); |
| 144 | } |
| 145 | |
| 146 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 147 | _mm_maskz_multishift_epi64_epi8(__mmask16 __M, __m128i __X, __m128i __Y) |
| 148 | { |
| 149 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
| 150 | (__v16qi)_mm_multishift_epi64_epi8(__X, __Y), |
| 151 | (__v16qi)_mm_setzero_si128()); |
| 152 | } |
| 153 | |
| 154 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 155 | _mm256_multishift_epi64_epi8(__m256i __X, __m256i __Y) |
| 156 | { |
| 157 | return (__m256i)__builtin_ia32_vpmultishiftqb256((__v32qi)__X, (__v32qi)__Y); |
| 158 | } |
| 159 | |
| 160 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 161 | _mm256_mask_multishift_epi64_epi8(__m256i __W, __mmask32 __M, __m256i __X, |
| 162 | __m256i __Y) |
| 163 | { |
| 164 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
| 165 | (__v32qi)_mm256_multishift_epi64_epi8(__X, __Y), |
| 166 | (__v32qi)__W); |
| 167 | } |
| 168 | |
| 169 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 170 | _mm256_maskz_multishift_epi64_epi8(__mmask32 __M, __m256i __X, __m256i __Y) |
| 171 | { |
| 172 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
| 173 | (__v32qi)_mm256_multishift_epi64_epi8(__X, __Y), |
| 174 | (__v32qi)_mm256_setzero_si256()); |
| 175 | } |
| 176 | |
| 177 | #undef __DEFAULT_FN_ATTRS128 |
| 178 | #undef __DEFAULT_FN_ATTRS256 |
| 179 | #endif |