Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line | 
|---|---|---|---|
| 14 | pmbaty | 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 | #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512cd"), __min_vector_width__(128))) | 
        ||
| 18 | #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512cd"), __min_vector_width__(256))) | 
        ||
| 19 | |||
| 20 | |||
| 21 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 22 | _mm_broadcastmb_epi64 (__mmask8 __A)  | 
        ||
| 23 | { | 
        ||
| 24 | return (__m128i) _mm_set1_epi64x((long long) __A);  | 
        ||
| 25 | } | 
        ||
| 26 | |||
| 27 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 28 | _mm256_broadcastmb_epi64 (__mmask8 __A)  | 
        ||
| 29 | { | 
        ||
| 30 | return (__m256i) _mm256_set1_epi64x((long long)__A);  | 
        ||
| 31 | } | 
        ||
| 32 | |||
| 33 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 34 | _mm_broadcastmw_epi32 (__mmask16 __A)  | 
        ||
| 35 | { | 
        ||
| 36 | return (__m128i) _mm_set1_epi32((int)__A);  | 
        ||
| 37 | } | 
        ||
| 38 | |||
| 39 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 40 | _mm256_broadcastmw_epi32 (__mmask16 __A)  | 
        ||
| 41 | { | 
        ||
| 42 | return (__m256i) _mm256_set1_epi32((int)__A);  | 
        ||
| 43 | } | 
        ||
| 44 | |||
| 45 | |||
| 46 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 47 | _mm_conflict_epi64 (__m128i __A)  | 
        ||
| 48 | { | 
        ||
| 49 | return (__m128i) __builtin_ia32_vpconflictdi_128 ((__v2di) __A);  | 
        ||
| 50 | } | 
        ||
| 51 | |||
| 52 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 53 | _mm_mask_conflict_epi64 (__m128i __W, __mmask8 __U, __m128i __A)  | 
        ||
| 54 | { | 
        ||
| 55 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,  | 
        ||
| 56 | (__v2di)_mm_conflict_epi64(__A),  | 
        ||
| 57 | (__v2di)__W);  | 
        ||
| 58 | } | 
        ||
| 59 | |||
| 60 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 61 | _mm_maskz_conflict_epi64 (__mmask8 __U, __m128i __A)  | 
        ||
| 62 | { | 
        ||
| 63 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,  | 
        ||
| 64 | (__v2di)_mm_conflict_epi64(__A),  | 
        ||
| 65 | (__v2di)_mm_setzero_si128());  | 
        ||
| 66 | } | 
        ||
| 67 | |||
| 68 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 69 | _mm256_conflict_epi64 (__m256i __A)  | 
        ||
| 70 | { | 
        ||
| 71 | return (__m256i) __builtin_ia32_vpconflictdi_256 ((__v4di) __A);  | 
        ||
| 72 | } | 
        ||
| 73 | |||
| 74 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 75 | _mm256_mask_conflict_epi64 (__m256i __W, __mmask8 __U, __m256i __A)  | 
        ||
| 76 | { | 
        ||
| 77 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,  | 
        ||
| 78 | (__v4di)_mm256_conflict_epi64(__A),  | 
        ||
| 79 | (__v4di)__W);  | 
        ||
| 80 | } | 
        ||
| 81 | |||
| 82 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 83 | _mm256_maskz_conflict_epi64 (__mmask8 __U, __m256i __A)  | 
        ||
| 84 | { | 
        ||
| 85 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,  | 
        ||
| 86 | (__v4di)_mm256_conflict_epi64(__A),  | 
        ||
| 87 | (__v4di)_mm256_setzero_si256());  | 
        ||
| 88 | } | 
        ||
| 89 | |||
| 90 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 91 | _mm_conflict_epi32 (__m128i __A)  | 
        ||
| 92 | { | 
        ||
| 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 | { | 
        ||
| 99 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,  | 
        ||
| 100 | (__v4si)_mm_conflict_epi32(__A),  | 
        ||
| 101 | (__v4si)__W);  | 
        ||
| 102 | } | 
        ||
| 103 | |||
| 104 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 105 | _mm_maskz_conflict_epi32 (__mmask8 __U, __m128i __A)  | 
        ||
| 106 | { | 
        ||
| 107 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,  | 
        ||
| 108 | (__v4si)_mm_conflict_epi32(__A),  | 
        ||
| 109 | (__v4si)_mm_setzero_si128());  | 
        ||
| 110 | } | 
        ||
| 111 | |||
| 112 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 113 | _mm256_conflict_epi32 (__m256i __A)  | 
        ||
| 114 | { | 
        ||
| 115 | return (__m256i) __builtin_ia32_vpconflictsi_256 ((__v8si) __A);  | 
        ||
| 116 | } | 
        ||
| 117 | |||
| 118 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 119 | _mm256_mask_conflict_epi32 (__m256i __W, __mmask8 __U, __m256i __A)  | 
        ||
| 120 | { | 
        ||
| 121 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,  | 
        ||
| 122 | (__v8si)_mm256_conflict_epi32(__A),  | 
        ||
| 123 | (__v8si)__W);  | 
        ||
| 124 | } | 
        ||
| 125 | |||
| 126 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 127 | _mm256_maskz_conflict_epi32 (__mmask8 __U, __m256i __A)  | 
        ||
| 128 | { | 
        ||
| 129 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,  | 
        ||
| 130 | (__v8si)_mm256_conflict_epi32(__A),  | 
        ||
| 131 | (__v8si)_mm256_setzero_si256());  | 
        ||
| 132 | } | 
        ||
| 133 | |||
| 134 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 135 | _mm_lzcnt_epi32 (__m128i __A)  | 
        ||
| 136 | { | 
        ||
| 137 | return (__m128i) __builtin_ia32_vplzcntd_128 ((__v4si) __A);  | 
        ||
| 138 | } | 
        ||
| 139 | |||
| 140 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 141 | _mm_mask_lzcnt_epi32 (__m128i __W, __mmask8 __U, __m128i __A)  | 
        ||
| 142 | { | 
        ||
| 143 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,  | 
        ||
| 144 | (__v4si)_mm_lzcnt_epi32(__A),  | 
        ||
| 145 | (__v4si)__W);  | 
        ||
| 146 | } | 
        ||
| 147 | |||
| 148 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 149 | _mm_maskz_lzcnt_epi32 (__mmask8 __U, __m128i __A)  | 
        ||
| 150 | { | 
        ||
| 151 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U,  | 
        ||
| 152 | (__v4si)_mm_lzcnt_epi32(__A),  | 
        ||
| 153 | (__v4si)_mm_setzero_si128());  | 
        ||
| 154 | } | 
        ||
| 155 | |||
| 156 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 157 | _mm256_lzcnt_epi32 (__m256i __A)  | 
        ||
| 158 | { | 
        ||
| 159 | return (__m256i) __builtin_ia32_vplzcntd_256 ((__v8si) __A);  | 
        ||
| 160 | } | 
        ||
| 161 | |||
| 162 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 163 | _mm256_mask_lzcnt_epi32 (__m256i __W, __mmask8 __U, __m256i __A)  | 
        ||
| 164 | { | 
        ||
| 165 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,  | 
        ||
| 166 | (__v8si)_mm256_lzcnt_epi32(__A),  | 
        ||
| 167 | (__v8si)__W);  | 
        ||
| 168 | } | 
        ||
| 169 | |||
| 170 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 171 | _mm256_maskz_lzcnt_epi32 (__mmask8 __U, __m256i __A)  | 
        ||
| 172 | { | 
        ||
| 173 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U,  | 
        ||
| 174 | (__v8si)_mm256_lzcnt_epi32(__A),  | 
        ||
| 175 | (__v8si)_mm256_setzero_si256());  | 
        ||
| 176 | } | 
        ||
| 177 | |||
| 178 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 179 | _mm_lzcnt_epi64 (__m128i __A)  | 
        ||
| 180 | { | 
        ||
| 181 | return (__m128i) __builtin_ia32_vplzcntq_128 ((__v2di) __A);  | 
        ||
| 182 | } | 
        ||
| 183 | |||
| 184 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 185 | _mm_mask_lzcnt_epi64 (__m128i __W, __mmask8 __U, __m128i __A)  | 
        ||
| 186 | { | 
        ||
| 187 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,  | 
        ||
| 188 | (__v2di)_mm_lzcnt_epi64(__A),  | 
        ||
| 189 | (__v2di)__W);  | 
        ||
| 190 | } | 
        ||
| 191 | |||
| 192 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 | 
        ||
| 193 | _mm_maskz_lzcnt_epi64 (__mmask8 __U, __m128i __A)  | 
        ||
| 194 | { | 
        ||
| 195 | return (__m128i)__builtin_ia32_selectq_128((__mmask8)__U,  | 
        ||
| 196 | (__v2di)_mm_lzcnt_epi64(__A),  | 
        ||
| 197 | (__v2di)_mm_setzero_si128());  | 
        ||
| 198 | } | 
        ||
| 199 | |||
| 200 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 201 | _mm256_lzcnt_epi64 (__m256i __A)  | 
        ||
| 202 | { | 
        ||
| 203 | return (__m256i) __builtin_ia32_vplzcntq_256 ((__v4di) __A);  | 
        ||
| 204 | } | 
        ||
| 205 | |||
| 206 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 207 | _mm256_mask_lzcnt_epi64 (__m256i __W, __mmask8 __U, __m256i __A)  | 
        ||
| 208 | { | 
        ||
| 209 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,  | 
        ||
| 210 | (__v4di)_mm256_lzcnt_epi64(__A),  | 
        ||
| 211 | (__v4di)__W);  | 
        ||
| 212 | } | 
        ||
| 213 | |||
| 214 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 | 
        ||
| 215 | _mm256_maskz_lzcnt_epi64 (__mmask8 __U, __m256i __A)  | 
        ||
| 216 | { | 
        ||
| 217 | return (__m256i)__builtin_ia32_selectq_256((__mmask8)__U,  | 
        ||
| 218 | (__v4di)_mm256_lzcnt_epi64(__A),  | 
        ||
| 219 | (__v4di)_mm256_setzero_si256());  | 
        ||
| 220 | } | 
        ||
| 221 | |||
| 222 | #undef __DEFAULT_FN_ATTRS128 | 
        ||
| 223 | #undef __DEFAULT_FN_ATTRS256 | 
        ||
| 224 | |||
| 225 | #endif /* __AVX512VLCDINTRIN_H */ |