| 1 | /*===---- avx512vlcdintrin.h - AVX512VL and AVX512CD 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 | #ifndef __IMMINTRIN_H |
| 10 | #error "Never use <avx512vlcdintrin.h> directly; include <immintrin.h> instead." |
| 11 | #endif |
| 12 | |
| 13 | #ifndef __AVX512VLCDINTRIN_H |
| 14 | #define __AVX512VLCDINTRIN_H |
| 15 | |
| 16 | /* Define the default attributes for the functions in this file. */ |
| 17 | #if defined(__cplusplus) && (__cplusplus >= 201103L) |
| 18 | #define __DEFAULT_FN_ATTRS128 \ |
| 19 | __attribute__((__always_inline__, __nodebug__, \ |
| 20 | __target__("avx512vl,avx512cd"), \ |
| 21 | __min_vector_width__(128))) constexpr |
| 22 | #define __DEFAULT_FN_ATTRS256 \ |
| 23 | __attribute__((__always_inline__, __nodebug__, \ |
| 24 | __target__("avx512vl,avx512cd"), \ |
| 25 | __min_vector_width__(256))) constexpr |
| 26 | #else |
| 27 | #define __DEFAULT_FN_ATTRS128 \ |
| 28 | __attribute__((__always_inline__, __nodebug__, \ |
| 29 | __target__("avx512vl,avx512cd"), __min_vector_width__(128))) |
| 30 | #define __DEFAULT_FN_ATTRS256 \ |
| 31 | __attribute__((__always_inline__, __nodebug__, \ |
| 32 | __target__("avx512vl,avx512cd"), __min_vector_width__(256))) |
| 33 | #endif |
| 34 | |
| 35 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 36 | _mm_broadcastmb_epi64(__mmask8 __A) { |
| 37 | return (__m128i) _mm_set1_epi64x((long long) __A); |
| 38 | } |
| 39 | |
| 40 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 41 | _mm256_broadcastmb_epi64(__mmask8 __A) { |
| 42 | return (__m256i)_mm256_set1_epi64x((long long)__A); |
| 43 | } |
| 44 | |
| 45 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 46 | _mm_broadcastmw_epi32(__mmask16 __A) { |
| 47 | return (__m128i) _mm_set1_epi32((int)__A); |
| 48 | } |
| 49 | |
| 50 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 51 | _mm256_broadcastmw_epi32(__mmask16 __A) { |
| 52 | return (__m256i) _mm256_set1_epi32((int)__A); |
| 53 | } |
| 54 | |
| 55 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 56 | _mm_conflict_epi64(__m128i __A) { |
| 57 | return (__m128i)__builtin_ia32_vpconflictdi_128((__v2di)__A); |
| 58 | } |
| 59 | |
| 60 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 61 | _mm_mask_conflict_epi64(__m128i __W, __mmask8 __U, __m128i __A) { |
| 62 | return (__m128i)__builtin_ia32_selectq_128( |
| 63 | (__mmask8)__U, (__v2di)_mm_conflict_epi64(__A), (__v2di)__W); |
| 64 | } |
| 65 | |
| 66 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 67 | _mm_maskz_conflict_epi64(__mmask8 __U, __m128i __A) { |
| 68 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U, |
| 69 | (__v2di)_mm_conflict_epi64(__A), |
| 70 | (__v2di)_mm_setzero_si128()); |
| 71 | } |
| 72 | |
| 73 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 74 | _mm256_conflict_epi64(__m256i __A) { |
| 75 | return (__m256i)__builtin_ia32_vpconflictdi_256((__v4di)__A); |
| 76 | } |
| 77 | |
| 78 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 79 | _mm256_mask_conflict_epi64(__m256i __W, __mmask8 __U, __m256i __A) { |
| 80 | return (__m256i)__builtin_ia32_selectq_256( |
| 81 | (__mmask8)__U, (__v4di)_mm256_conflict_epi64(__A), (__v4di)__W); |
| 82 | } |
| 83 | |
| 84 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 85 | _mm256_maskz_conflict_epi64(__mmask8 __U, __m256i __A) { |
| 86 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 87 | (__v4di)_mm256_conflict_epi64(__A), |
| 88 | (__v4di)_mm256_setzero_si256()); |
| 89 | } |
| 90 | |
| 91 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 92 | _mm_conflict_epi32(__m128i __A) { |
| 93 | return (__m128i)__builtin_ia32_vpconflictsi_128((__v4si)__A); |
| 94 | } |
| 95 | |
| 96 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 97 | _mm_mask_conflict_epi32(__m128i __W, __mmask8 __U, __m128i __A) { |
| 98 | return (__m128i)__builtin_ia32_selectd_128( |
| 99 | (__mmask8)__U, (__v4si)_mm_conflict_epi32(__A), (__v4si)__W); |
| 100 | } |
| 101 | |
| 102 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 103 | _mm_maskz_conflict_epi32(__mmask8 __U, __m128i __A) { |
| 104 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
| 105 | (__v4si)_mm_conflict_epi32(__A), |
| 106 | (__v4si)_mm_setzero_si128()); |
| 107 | } |
| 108 | |
| 109 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 110 | _mm256_conflict_epi32(__m256i __A) { |
| 111 | return (__m256i)__builtin_ia32_vpconflictsi_256((__v8si)__A); |
| 112 | } |
| 113 | |
| 114 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 115 | _mm256_mask_conflict_epi32(__m256i __W, __mmask8 __U, __m256i __A) { |
| 116 | return (__m256i)__builtin_ia32_selectd_256( |
| 117 | (__mmask8)__U, (__v8si)_mm256_conflict_epi32(__A), (__v8si)__W); |
| 118 | } |
| 119 | |
| 120 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 121 | _mm256_maskz_conflict_epi32(__mmask8 __U, __m256i __A) { |
| 122 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 123 | (__v8si)_mm256_conflict_epi32(__A), |
| 124 | (__v8si)_mm256_setzero_si256()); |
| 125 | } |
| 126 | |
| 127 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_lzcnt_epi32(__m128i __A) { |
| 128 | return (__m128i)__builtin_elementwise_clzg((__v4si)__A, |
| 129 | (__v4si)_mm_set1_epi32(32)); |
| 130 | } |
| 131 | |
| 132 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 133 | _mm_mask_lzcnt_epi32(__m128i __W, __mmask8 __U, __m128i __A) { |
| 134 | return (__m128i)__builtin_ia32_selectd_128( |
| 135 | (__mmask8)__U, (__v4si)_mm_lzcnt_epi32(__A), (__v4si)__W); |
| 136 | } |
| 137 | |
| 138 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 139 | _mm_maskz_lzcnt_epi32(__mmask8 __U, __m128i __A) { |
| 140 | return (__m128i)__builtin_ia32_selectd_128( |
| 141 | (__mmask8)__U, (__v4si)_mm_lzcnt_epi32(__A), (__v4si)_mm_setzero_si128()); |
| 142 | } |
| 143 | |
| 144 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 145 | _mm256_lzcnt_epi32(__m256i __A) { |
| 146 | return (__m256i)__builtin_elementwise_clzg((__v8si)__A, |
| 147 | (__v8si)_mm256_set1_epi32(32)); |
| 148 | } |
| 149 | |
| 150 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 151 | _mm256_mask_lzcnt_epi32(__m256i __W, __mmask8 __U, __m256i __A) { |
| 152 | return (__m256i)__builtin_ia32_selectd_256( |
| 153 | (__mmask8)__U, (__v8si)_mm256_lzcnt_epi32(__A), (__v8si)__W); |
| 154 | } |
| 155 | |
| 156 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 157 | _mm256_maskz_lzcnt_epi32(__mmask8 __U, __m256i __A) { |
| 158 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
| 159 | (__v8si)_mm256_lzcnt_epi32(__A), |
| 160 | (__v8si)_mm256_setzero_si256()); |
| 161 | } |
| 162 | |
| 163 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 _mm_lzcnt_epi64(__m128i __A) { |
| 164 | return (__m128i)__builtin_elementwise_clzg( |
| 165 | (__v2di)__A, (__v2di)_mm_set1_epi64x((long long)64)); |
| 166 | } |
| 167 | |
| 168 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 169 | _mm_mask_lzcnt_epi64(__m128i __W, __mmask8 __U, __m128i __A) { |
| 170 | return (__m128i)__builtin_ia32_selectq_128( |
| 171 | (__mmask8)__U, (__v2di)_mm_lzcnt_epi64(__A), (__v2di)__W); |
| 172 | } |
| 173 | |
| 174 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
| 175 | _mm_maskz_lzcnt_epi64(__mmask8 __U, __m128i __A) { |
| 176 | return (__m128i)__builtin_ia32_selectq_128( |
| 177 | (__mmask8)__U, (__v2di)_mm_lzcnt_epi64(__A), (__v2di)_mm_setzero_si128()); |
| 178 | } |
| 179 | |
| 180 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 181 | _mm256_lzcnt_epi64(__m256i __A) { |
| 182 | return (__m256i)__builtin_elementwise_clzg( |
| 183 | (__v4di)__A, (__v4di)_mm256_set1_epi64x((long long)64)); |
| 184 | } |
| 185 | |
| 186 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 187 | _mm256_mask_lzcnt_epi64(__m256i __W, __mmask8 __U, __m256i __A) { |
| 188 | return (__m256i)__builtin_ia32_selectq_256( |
| 189 | (__mmask8)__U, (__v4di)_mm256_lzcnt_epi64(__A), (__v4di)__W); |
| 190 | } |
| 191 | |
| 192 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
| 193 | _mm256_maskz_lzcnt_epi64(__mmask8 __U, __m256i __A) { |
| 194 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U, |
| 195 | (__v4di)_mm256_lzcnt_epi64(__A), |
| 196 | (__v4di)_mm256_setzero_si256()); |
| 197 | } |
| 198 | |
| 199 | #undef __DEFAULT_FN_ATTRS128 |
| 200 | #undef __DEFAULT_FN_ATTRS256 |
| 201 | |
| 202 | #endif /* __AVX512VLCDINTRIN_H */ |