Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | /*===---- avx512vlbwintrin.h - AVX512VL and AVX512BW 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 | |||
| 10 | #ifndef __IMMINTRIN_H |
||
| 11 | #error "Never use <avx512vlbwintrin.h> directly; include <immintrin.h> instead." |
||
| 12 | #endif |
||
| 13 | |||
| 14 | #ifndef __AVX512VLBWINTRIN_H |
||
| 15 | #define __AVX512VLBWINTRIN_H |
||
| 16 | |||
| 17 | /* Define the default attributes for the functions in this file. */ |
||
| 18 | #define __DEFAULT_FN_ATTRS128 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bw"), __min_vector_width__(128))) |
||
| 19 | #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("avx512vl,avx512bw"), __min_vector_width__(256))) |
||
| 20 | |||
| 21 | /* Integer compare */ |
||
| 22 | |||
| 23 | #define _mm_cmp_epi8_mask(a, b, p) \ |
||
| 24 | ((__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \ |
||
| 25 | (__v16qi)(__m128i)(b), (int)(p), \ |
||
| 26 | (__mmask16)-1)) |
||
| 27 | |||
| 28 | #define _mm_mask_cmp_epi8_mask(m, a, b, p) \ |
||
| 29 | ((__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \ |
||
| 30 | (__v16qi)(__m128i)(b), (int)(p), \ |
||
| 31 | (__mmask16)(m))) |
||
| 32 | |||
| 33 | #define _mm_cmp_epu8_mask(a, b, p) \ |
||
| 34 | ((__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \ |
||
| 35 | (__v16qi)(__m128i)(b), (int)(p), \ |
||
| 36 | (__mmask16)-1)) |
||
| 37 | |||
| 38 | #define _mm_mask_cmp_epu8_mask(m, a, b, p) \ |
||
| 39 | ((__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \ |
||
| 40 | (__v16qi)(__m128i)(b), (int)(p), \ |
||
| 41 | (__mmask16)(m))) |
||
| 42 | |||
| 43 | #define _mm256_cmp_epi8_mask(a, b, p) \ |
||
| 44 | ((__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \ |
||
| 45 | (__v32qi)(__m256i)(b), (int)(p), \ |
||
| 46 | (__mmask32)-1)) |
||
| 47 | |||
| 48 | #define _mm256_mask_cmp_epi8_mask(m, a, b, p) \ |
||
| 49 | ((__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \ |
||
| 50 | (__v32qi)(__m256i)(b), (int)(p), \ |
||
| 51 | (__mmask32)(m))) |
||
| 52 | |||
| 53 | #define _mm256_cmp_epu8_mask(a, b, p) \ |
||
| 54 | ((__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \ |
||
| 55 | (__v32qi)(__m256i)(b), (int)(p), \ |
||
| 56 | (__mmask32)-1)) |
||
| 57 | |||
| 58 | #define _mm256_mask_cmp_epu8_mask(m, a, b, p) \ |
||
| 59 | ((__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \ |
||
| 60 | (__v32qi)(__m256i)(b), (int)(p), \ |
||
| 61 | (__mmask32)(m))) |
||
| 62 | |||
| 63 | #define _mm_cmp_epi16_mask(a, b, p) \ |
||
| 64 | ((__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \ |
||
| 65 | (__v8hi)(__m128i)(b), (int)(p), \ |
||
| 66 | (__mmask8)-1)) |
||
| 67 | |||
| 68 | #define _mm_mask_cmp_epi16_mask(m, a, b, p) \ |
||
| 69 | ((__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \ |
||
| 70 | (__v8hi)(__m128i)(b), (int)(p), \ |
||
| 71 | (__mmask8)(m))) |
||
| 72 | |||
| 73 | #define _mm_cmp_epu16_mask(a, b, p) \ |
||
| 74 | ((__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \ |
||
| 75 | (__v8hi)(__m128i)(b), (int)(p), \ |
||
| 76 | (__mmask8)-1)) |
||
| 77 | |||
| 78 | #define _mm_mask_cmp_epu16_mask(m, a, b, p) \ |
||
| 79 | ((__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \ |
||
| 80 | (__v8hi)(__m128i)(b), (int)(p), \ |
||
| 81 | (__mmask8)(m))) |
||
| 82 | |||
| 83 | #define _mm256_cmp_epi16_mask(a, b, p) \ |
||
| 84 | ((__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \ |
||
| 85 | (__v16hi)(__m256i)(b), (int)(p), \ |
||
| 86 | (__mmask16)-1)) |
||
| 87 | |||
| 88 | #define _mm256_mask_cmp_epi16_mask(m, a, b, p) \ |
||
| 89 | ((__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \ |
||
| 90 | (__v16hi)(__m256i)(b), (int)(p), \ |
||
| 91 | (__mmask16)(m))) |
||
| 92 | |||
| 93 | #define _mm256_cmp_epu16_mask(a, b, p) \ |
||
| 94 | ((__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \ |
||
| 95 | (__v16hi)(__m256i)(b), (int)(p), \ |
||
| 96 | (__mmask16)-1)) |
||
| 97 | |||
| 98 | #define _mm256_mask_cmp_epu16_mask(m, a, b, p) \ |
||
| 99 | ((__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \ |
||
| 100 | (__v16hi)(__m256i)(b), (int)(p), \ |
||
| 101 | (__mmask16)(m))) |
||
| 102 | |||
| 103 | #define _mm_cmpeq_epi8_mask(A, B) \ |
||
| 104 | _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ) |
||
| 105 | #define _mm_mask_cmpeq_epi8_mask(k, A, B) \ |
||
| 106 | _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 107 | #define _mm_cmpge_epi8_mask(A, B) \ |
||
| 108 | _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_GE) |
||
| 109 | #define _mm_mask_cmpge_epi8_mask(k, A, B) \ |
||
| 110 | _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 111 | #define _mm_cmpgt_epi8_mask(A, B) \ |
||
| 112 | _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_GT) |
||
| 113 | #define _mm_mask_cmpgt_epi8_mask(k, A, B) \ |
||
| 114 | _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 115 | #define _mm_cmple_epi8_mask(A, B) \ |
||
| 116 | _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_LE) |
||
| 117 | #define _mm_mask_cmple_epi8_mask(k, A, B) \ |
||
| 118 | _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 119 | #define _mm_cmplt_epi8_mask(A, B) \ |
||
| 120 | _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_LT) |
||
| 121 | #define _mm_mask_cmplt_epi8_mask(k, A, B) \ |
||
| 122 | _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 123 | #define _mm_cmpneq_epi8_mask(A, B) \ |
||
| 124 | _mm_cmp_epi8_mask((A), (B), _MM_CMPINT_NE) |
||
| 125 | #define _mm_mask_cmpneq_epi8_mask(k, A, B) \ |
||
| 126 | _mm_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 127 | |||
| 128 | #define _mm256_cmpeq_epi8_mask(A, B) \ |
||
| 129 | _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_EQ) |
||
| 130 | #define _mm256_mask_cmpeq_epi8_mask(k, A, B) \ |
||
| 131 | _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 132 | #define _mm256_cmpge_epi8_mask(A, B) \ |
||
| 133 | _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_GE) |
||
| 134 | #define _mm256_mask_cmpge_epi8_mask(k, A, B) \ |
||
| 135 | _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 136 | #define _mm256_cmpgt_epi8_mask(A, B) \ |
||
| 137 | _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_GT) |
||
| 138 | #define _mm256_mask_cmpgt_epi8_mask(k, A, B) \ |
||
| 139 | _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 140 | #define _mm256_cmple_epi8_mask(A, B) \ |
||
| 141 | _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_LE) |
||
| 142 | #define _mm256_mask_cmple_epi8_mask(k, A, B) \ |
||
| 143 | _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 144 | #define _mm256_cmplt_epi8_mask(A, B) \ |
||
| 145 | _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_LT) |
||
| 146 | #define _mm256_mask_cmplt_epi8_mask(k, A, B) \ |
||
| 147 | _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 148 | #define _mm256_cmpneq_epi8_mask(A, B) \ |
||
| 149 | _mm256_cmp_epi8_mask((A), (B), _MM_CMPINT_NE) |
||
| 150 | #define _mm256_mask_cmpneq_epi8_mask(k, A, B) \ |
||
| 151 | _mm256_mask_cmp_epi8_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 152 | |||
| 153 | #define _mm_cmpeq_epu8_mask(A, B) \ |
||
| 154 | _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ) |
||
| 155 | #define _mm_mask_cmpeq_epu8_mask(k, A, B) \ |
||
| 156 | _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 157 | #define _mm_cmpge_epu8_mask(A, B) \ |
||
| 158 | _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_GE) |
||
| 159 | #define _mm_mask_cmpge_epu8_mask(k, A, B) \ |
||
| 160 | _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 161 | #define _mm_cmpgt_epu8_mask(A, B) \ |
||
| 162 | _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_GT) |
||
| 163 | #define _mm_mask_cmpgt_epu8_mask(k, A, B) \ |
||
| 164 | _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 165 | #define _mm_cmple_epu8_mask(A, B) \ |
||
| 166 | _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_LE) |
||
| 167 | #define _mm_mask_cmple_epu8_mask(k, A, B) \ |
||
| 168 | _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 169 | #define _mm_cmplt_epu8_mask(A, B) \ |
||
| 170 | _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_LT) |
||
| 171 | #define _mm_mask_cmplt_epu8_mask(k, A, B) \ |
||
| 172 | _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 173 | #define _mm_cmpneq_epu8_mask(A, B) \ |
||
| 174 | _mm_cmp_epu8_mask((A), (B), _MM_CMPINT_NE) |
||
| 175 | #define _mm_mask_cmpneq_epu8_mask(k, A, B) \ |
||
| 176 | _mm_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 177 | |||
| 178 | #define _mm256_cmpeq_epu8_mask(A, B) \ |
||
| 179 | _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_EQ) |
||
| 180 | #define _mm256_mask_cmpeq_epu8_mask(k, A, B) \ |
||
| 181 | _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 182 | #define _mm256_cmpge_epu8_mask(A, B) \ |
||
| 183 | _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_GE) |
||
| 184 | #define _mm256_mask_cmpge_epu8_mask(k, A, B) \ |
||
| 185 | _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 186 | #define _mm256_cmpgt_epu8_mask(A, B) \ |
||
| 187 | _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_GT) |
||
| 188 | #define _mm256_mask_cmpgt_epu8_mask(k, A, B) \ |
||
| 189 | _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 190 | #define _mm256_cmple_epu8_mask(A, B) \ |
||
| 191 | _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_LE) |
||
| 192 | #define _mm256_mask_cmple_epu8_mask(k, A, B) \ |
||
| 193 | _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 194 | #define _mm256_cmplt_epu8_mask(A, B) \ |
||
| 195 | _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_LT) |
||
| 196 | #define _mm256_mask_cmplt_epu8_mask(k, A, B) \ |
||
| 197 | _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 198 | #define _mm256_cmpneq_epu8_mask(A, B) \ |
||
| 199 | _mm256_cmp_epu8_mask((A), (B), _MM_CMPINT_NE) |
||
| 200 | #define _mm256_mask_cmpneq_epu8_mask(k, A, B) \ |
||
| 201 | _mm256_mask_cmp_epu8_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 202 | |||
| 203 | #define _mm_cmpeq_epi16_mask(A, B) \ |
||
| 204 | _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ) |
||
| 205 | #define _mm_mask_cmpeq_epi16_mask(k, A, B) \ |
||
| 206 | _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 207 | #define _mm_cmpge_epi16_mask(A, B) \ |
||
| 208 | _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_GE) |
||
| 209 | #define _mm_mask_cmpge_epi16_mask(k, A, B) \ |
||
| 210 | _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 211 | #define _mm_cmpgt_epi16_mask(A, B) \ |
||
| 212 | _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_GT) |
||
| 213 | #define _mm_mask_cmpgt_epi16_mask(k, A, B) \ |
||
| 214 | _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 215 | #define _mm_cmple_epi16_mask(A, B) \ |
||
| 216 | _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_LE) |
||
| 217 | #define _mm_mask_cmple_epi16_mask(k, A, B) \ |
||
| 218 | _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 219 | #define _mm_cmplt_epi16_mask(A, B) \ |
||
| 220 | _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_LT) |
||
| 221 | #define _mm_mask_cmplt_epi16_mask(k, A, B) \ |
||
| 222 | _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 223 | #define _mm_cmpneq_epi16_mask(A, B) \ |
||
| 224 | _mm_cmp_epi16_mask((A), (B), _MM_CMPINT_NE) |
||
| 225 | #define _mm_mask_cmpneq_epi16_mask(k, A, B) \ |
||
| 226 | _mm_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 227 | |||
| 228 | #define _mm256_cmpeq_epi16_mask(A, B) \ |
||
| 229 | _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_EQ) |
||
| 230 | #define _mm256_mask_cmpeq_epi16_mask(k, A, B) \ |
||
| 231 | _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 232 | #define _mm256_cmpge_epi16_mask(A, B) \ |
||
| 233 | _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_GE) |
||
| 234 | #define _mm256_mask_cmpge_epi16_mask(k, A, B) \ |
||
| 235 | _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 236 | #define _mm256_cmpgt_epi16_mask(A, B) \ |
||
| 237 | _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_GT) |
||
| 238 | #define _mm256_mask_cmpgt_epi16_mask(k, A, B) \ |
||
| 239 | _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 240 | #define _mm256_cmple_epi16_mask(A, B) \ |
||
| 241 | _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_LE) |
||
| 242 | #define _mm256_mask_cmple_epi16_mask(k, A, B) \ |
||
| 243 | _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 244 | #define _mm256_cmplt_epi16_mask(A, B) \ |
||
| 245 | _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_LT) |
||
| 246 | #define _mm256_mask_cmplt_epi16_mask(k, A, B) \ |
||
| 247 | _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 248 | #define _mm256_cmpneq_epi16_mask(A, B) \ |
||
| 249 | _mm256_cmp_epi16_mask((A), (B), _MM_CMPINT_NE) |
||
| 250 | #define _mm256_mask_cmpneq_epi16_mask(k, A, B) \ |
||
| 251 | _mm256_mask_cmp_epi16_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 252 | |||
| 253 | #define _mm_cmpeq_epu16_mask(A, B) \ |
||
| 254 | _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ) |
||
| 255 | #define _mm_mask_cmpeq_epu16_mask(k, A, B) \ |
||
| 256 | _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 257 | #define _mm_cmpge_epu16_mask(A, B) \ |
||
| 258 | _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_GE) |
||
| 259 | #define _mm_mask_cmpge_epu16_mask(k, A, B) \ |
||
| 260 | _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 261 | #define _mm_cmpgt_epu16_mask(A, B) \ |
||
| 262 | _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_GT) |
||
| 263 | #define _mm_mask_cmpgt_epu16_mask(k, A, B) \ |
||
| 264 | _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 265 | #define _mm_cmple_epu16_mask(A, B) \ |
||
| 266 | _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_LE) |
||
| 267 | #define _mm_mask_cmple_epu16_mask(k, A, B) \ |
||
| 268 | _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 269 | #define _mm_cmplt_epu16_mask(A, B) \ |
||
| 270 | _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_LT) |
||
| 271 | #define _mm_mask_cmplt_epu16_mask(k, A, B) \ |
||
| 272 | _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 273 | #define _mm_cmpneq_epu16_mask(A, B) \ |
||
| 274 | _mm_cmp_epu16_mask((A), (B), _MM_CMPINT_NE) |
||
| 275 | #define _mm_mask_cmpneq_epu16_mask(k, A, B) \ |
||
| 276 | _mm_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 277 | |||
| 278 | #define _mm256_cmpeq_epu16_mask(A, B) \ |
||
| 279 | _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_EQ) |
||
| 280 | #define _mm256_mask_cmpeq_epu16_mask(k, A, B) \ |
||
| 281 | _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_EQ) |
||
| 282 | #define _mm256_cmpge_epu16_mask(A, B) \ |
||
| 283 | _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_GE) |
||
| 284 | #define _mm256_mask_cmpge_epu16_mask(k, A, B) \ |
||
| 285 | _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GE) |
||
| 286 | #define _mm256_cmpgt_epu16_mask(A, B) \ |
||
| 287 | _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_GT) |
||
| 288 | #define _mm256_mask_cmpgt_epu16_mask(k, A, B) \ |
||
| 289 | _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_GT) |
||
| 290 | #define _mm256_cmple_epu16_mask(A, B) \ |
||
| 291 | _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_LE) |
||
| 292 | #define _mm256_mask_cmple_epu16_mask(k, A, B) \ |
||
| 293 | _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LE) |
||
| 294 | #define _mm256_cmplt_epu16_mask(A, B) \ |
||
| 295 | _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_LT) |
||
| 296 | #define _mm256_mask_cmplt_epu16_mask(k, A, B) \ |
||
| 297 | _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_LT) |
||
| 298 | #define _mm256_cmpneq_epu16_mask(A, B) \ |
||
| 299 | _mm256_cmp_epu16_mask((A), (B), _MM_CMPINT_NE) |
||
| 300 | #define _mm256_mask_cmpneq_epu16_mask(k, A, B) \ |
||
| 301 | _mm256_mask_cmp_epu16_mask((k), (A), (B), _MM_CMPINT_NE) |
||
| 302 | |||
| 303 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 304 | _mm256_mask_add_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B){ |
||
| 305 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 306 | (__v32qi)_mm256_add_epi8(__A, __B), |
||
| 307 | (__v32qi)__W); |
||
| 308 | } |
||
| 309 | |||
| 310 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 311 | _mm256_maskz_add_epi8(__mmask32 __U, __m256i __A, __m256i __B) { |
||
| 312 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 313 | (__v32qi)_mm256_add_epi8(__A, __B), |
||
| 314 | (__v32qi)_mm256_setzero_si256()); |
||
| 315 | } |
||
| 316 | |||
| 317 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 318 | _mm256_mask_add_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 319 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 320 | (__v16hi)_mm256_add_epi16(__A, __B), |
||
| 321 | (__v16hi)__W); |
||
| 322 | } |
||
| 323 | |||
| 324 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 325 | _mm256_maskz_add_epi16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 326 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 327 | (__v16hi)_mm256_add_epi16(__A, __B), |
||
| 328 | (__v16hi)_mm256_setzero_si256()); |
||
| 329 | } |
||
| 330 | |||
| 331 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 332 | _mm256_mask_sub_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) { |
||
| 333 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 334 | (__v32qi)_mm256_sub_epi8(__A, __B), |
||
| 335 | (__v32qi)__W); |
||
| 336 | } |
||
| 337 | |||
| 338 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 339 | _mm256_maskz_sub_epi8(__mmask32 __U, __m256i __A, __m256i __B) { |
||
| 340 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 341 | (__v32qi)_mm256_sub_epi8(__A, __B), |
||
| 342 | (__v32qi)_mm256_setzero_si256()); |
||
| 343 | } |
||
| 344 | |||
| 345 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 346 | _mm256_mask_sub_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 347 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 348 | (__v16hi)_mm256_sub_epi16(__A, __B), |
||
| 349 | (__v16hi)__W); |
||
| 350 | } |
||
| 351 | |||
| 352 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 353 | _mm256_maskz_sub_epi16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 354 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 355 | (__v16hi)_mm256_sub_epi16(__A, __B), |
||
| 356 | (__v16hi)_mm256_setzero_si256()); |
||
| 357 | } |
||
| 358 | |||
| 359 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 360 | _mm_mask_add_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) { |
||
| 361 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 362 | (__v16qi)_mm_add_epi8(__A, __B), |
||
| 363 | (__v16qi)__W); |
||
| 364 | } |
||
| 365 | |||
| 366 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 367 | _mm_maskz_add_epi8(__mmask16 __U, __m128i __A, __m128i __B) { |
||
| 368 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 369 | (__v16qi)_mm_add_epi8(__A, __B), |
||
| 370 | (__v16qi)_mm_setzero_si128()); |
||
| 371 | } |
||
| 372 | |||
| 373 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 374 | _mm_mask_add_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 375 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 376 | (__v8hi)_mm_add_epi16(__A, __B), |
||
| 377 | (__v8hi)__W); |
||
| 378 | } |
||
| 379 | |||
| 380 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 381 | _mm_maskz_add_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 382 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 383 | (__v8hi)_mm_add_epi16(__A, __B), |
||
| 384 | (__v8hi)_mm_setzero_si128()); |
||
| 385 | } |
||
| 386 | |||
| 387 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 388 | _mm_mask_sub_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) { |
||
| 389 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 390 | (__v16qi)_mm_sub_epi8(__A, __B), |
||
| 391 | (__v16qi)__W); |
||
| 392 | } |
||
| 393 | |||
| 394 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 395 | _mm_maskz_sub_epi8(__mmask16 __U, __m128i __A, __m128i __B) { |
||
| 396 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 397 | (__v16qi)_mm_sub_epi8(__A, __B), |
||
| 398 | (__v16qi)_mm_setzero_si128()); |
||
| 399 | } |
||
| 400 | |||
| 401 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 402 | _mm_mask_sub_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 403 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 404 | (__v8hi)_mm_sub_epi16(__A, __B), |
||
| 405 | (__v8hi)__W); |
||
| 406 | } |
||
| 407 | |||
| 408 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 409 | _mm_maskz_sub_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 410 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 411 | (__v8hi)_mm_sub_epi16(__A, __B), |
||
| 412 | (__v8hi)_mm_setzero_si128()); |
||
| 413 | } |
||
| 414 | |||
| 415 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 416 | _mm256_mask_mullo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 417 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 418 | (__v16hi)_mm256_mullo_epi16(__A, __B), |
||
| 419 | (__v16hi)__W); |
||
| 420 | } |
||
| 421 | |||
| 422 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 423 | _mm256_maskz_mullo_epi16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 424 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 425 | (__v16hi)_mm256_mullo_epi16(__A, __B), |
||
| 426 | (__v16hi)_mm256_setzero_si256()); |
||
| 427 | } |
||
| 428 | |||
| 429 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 430 | _mm_mask_mullo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 431 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 432 | (__v8hi)_mm_mullo_epi16(__A, __B), |
||
| 433 | (__v8hi)__W); |
||
| 434 | } |
||
| 435 | |||
| 436 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 437 | _mm_maskz_mullo_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 438 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 439 | (__v8hi)_mm_mullo_epi16(__A, __B), |
||
| 440 | (__v8hi)_mm_setzero_si128()); |
||
| 441 | } |
||
| 442 | |||
| 443 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 444 | _mm_mask_blend_epi8 (__mmask16 __U, __m128i __A, __m128i __W) |
||
| 445 | { |
||
| 446 | return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U, |
||
| 447 | (__v16qi) __W, |
||
| 448 | (__v16qi) __A); |
||
| 449 | } |
||
| 450 | |||
| 451 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 452 | _mm256_mask_blend_epi8 (__mmask32 __U, __m256i __A, __m256i __W) |
||
| 453 | { |
||
| 454 | return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U, |
||
| 455 | (__v32qi) __W, |
||
| 456 | (__v32qi) __A); |
||
| 457 | } |
||
| 458 | |||
| 459 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 460 | _mm_mask_blend_epi16 (__mmask8 __U, __m128i __A, __m128i __W) |
||
| 461 | { |
||
| 462 | return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U, |
||
| 463 | (__v8hi) __W, |
||
| 464 | (__v8hi) __A); |
||
| 465 | } |
||
| 466 | |||
| 467 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 468 | _mm256_mask_blend_epi16 (__mmask16 __U, __m256i __A, __m256i __W) |
||
| 469 | { |
||
| 470 | return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U, |
||
| 471 | (__v16hi) __W, |
||
| 472 | (__v16hi) __A); |
||
| 473 | } |
||
| 474 | |||
| 475 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 476 | _mm_mask_abs_epi8(__m128i __W, __mmask16 __U, __m128i __A) |
||
| 477 | { |
||
| 478 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 479 | (__v16qi)_mm_abs_epi8(__A), |
||
| 480 | (__v16qi)__W); |
||
| 481 | } |
||
| 482 | |||
| 483 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 484 | _mm_maskz_abs_epi8(__mmask16 __U, __m128i __A) |
||
| 485 | { |
||
| 486 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 487 | (__v16qi)_mm_abs_epi8(__A), |
||
| 488 | (__v16qi)_mm_setzero_si128()); |
||
| 489 | } |
||
| 490 | |||
| 491 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 492 | _mm256_mask_abs_epi8(__m256i __W, __mmask32 __U, __m256i __A) |
||
| 493 | { |
||
| 494 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 495 | (__v32qi)_mm256_abs_epi8(__A), |
||
| 496 | (__v32qi)__W); |
||
| 497 | } |
||
| 498 | |||
| 499 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 500 | _mm256_maskz_abs_epi8 (__mmask32 __U, __m256i __A) |
||
| 501 | { |
||
| 502 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 503 | (__v32qi)_mm256_abs_epi8(__A), |
||
| 504 | (__v32qi)_mm256_setzero_si256()); |
||
| 505 | } |
||
| 506 | |||
| 507 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 508 | _mm_mask_abs_epi16(__m128i __W, __mmask8 __U, __m128i __A) |
||
| 509 | { |
||
| 510 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 511 | (__v8hi)_mm_abs_epi16(__A), |
||
| 512 | (__v8hi)__W); |
||
| 513 | } |
||
| 514 | |||
| 515 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 516 | _mm_maskz_abs_epi16(__mmask8 __U, __m128i __A) |
||
| 517 | { |
||
| 518 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 519 | (__v8hi)_mm_abs_epi16(__A), |
||
| 520 | (__v8hi)_mm_setzero_si128()); |
||
| 521 | } |
||
| 522 | |||
| 523 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 524 | _mm256_mask_abs_epi16(__m256i __W, __mmask16 __U, __m256i __A) |
||
| 525 | { |
||
| 526 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 527 | (__v16hi)_mm256_abs_epi16(__A), |
||
| 528 | (__v16hi)__W); |
||
| 529 | } |
||
| 530 | |||
| 531 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 532 | _mm256_maskz_abs_epi16(__mmask16 __U, __m256i __A) |
||
| 533 | { |
||
| 534 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 535 | (__v16hi)_mm256_abs_epi16(__A), |
||
| 536 | (__v16hi)_mm256_setzero_si256()); |
||
| 537 | } |
||
| 538 | |||
| 539 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 540 | _mm_maskz_packs_epi32(__mmask8 __M, __m128i __A, __m128i __B) { |
||
| 541 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 542 | (__v8hi)_mm_packs_epi32(__A, __B), |
||
| 543 | (__v8hi)_mm_setzero_si128()); |
||
| 544 | } |
||
| 545 | |||
| 546 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 547 | _mm_mask_packs_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
||
| 548 | { |
||
| 549 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 550 | (__v8hi)_mm_packs_epi32(__A, __B), |
||
| 551 | (__v8hi)__W); |
||
| 552 | } |
||
| 553 | |||
| 554 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 555 | _mm256_maskz_packs_epi32(__mmask16 __M, __m256i __A, __m256i __B) |
||
| 556 | { |
||
| 557 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 558 | (__v16hi)_mm256_packs_epi32(__A, __B), |
||
| 559 | (__v16hi)_mm256_setzero_si256()); |
||
| 560 | } |
||
| 561 | |||
| 562 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 563 | _mm256_mask_packs_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) |
||
| 564 | { |
||
| 565 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 566 | (__v16hi)_mm256_packs_epi32(__A, __B), |
||
| 567 | (__v16hi)__W); |
||
| 568 | } |
||
| 569 | |||
| 570 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 571 | _mm_maskz_packs_epi16(__mmask16 __M, __m128i __A, __m128i __B) |
||
| 572 | { |
||
| 573 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 574 | (__v16qi)_mm_packs_epi16(__A, __B), |
||
| 575 | (__v16qi)_mm_setzero_si128()); |
||
| 576 | } |
||
| 577 | |||
| 578 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 579 | _mm_mask_packs_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) |
||
| 580 | { |
||
| 581 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 582 | (__v16qi)_mm_packs_epi16(__A, __B), |
||
| 583 | (__v16qi)__W); |
||
| 584 | } |
||
| 585 | |||
| 586 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 587 | _mm256_maskz_packs_epi16(__mmask32 __M, __m256i __A, __m256i __B) |
||
| 588 | { |
||
| 589 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 590 | (__v32qi)_mm256_packs_epi16(__A, __B), |
||
| 591 | (__v32qi)_mm256_setzero_si256()); |
||
| 592 | } |
||
| 593 | |||
| 594 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 595 | _mm256_mask_packs_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) |
||
| 596 | { |
||
| 597 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 598 | (__v32qi)_mm256_packs_epi16(__A, __B), |
||
| 599 | (__v32qi)__W); |
||
| 600 | } |
||
| 601 | |||
| 602 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 603 | _mm_maskz_packus_epi32(__mmask8 __M, __m128i __A, __m128i __B) |
||
| 604 | { |
||
| 605 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 606 | (__v8hi)_mm_packus_epi32(__A, __B), |
||
| 607 | (__v8hi)_mm_setzero_si128()); |
||
| 608 | } |
||
| 609 | |||
| 610 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 611 | _mm_mask_packus_epi32(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
||
| 612 | { |
||
| 613 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 614 | (__v8hi)_mm_packus_epi32(__A, __B), |
||
| 615 | (__v8hi)__W); |
||
| 616 | } |
||
| 617 | |||
| 618 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 619 | _mm256_maskz_packus_epi32(__mmask16 __M, __m256i __A, __m256i __B) |
||
| 620 | { |
||
| 621 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 622 | (__v16hi)_mm256_packus_epi32(__A, __B), |
||
| 623 | (__v16hi)_mm256_setzero_si256()); |
||
| 624 | } |
||
| 625 | |||
| 626 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 627 | _mm256_mask_packus_epi32(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) |
||
| 628 | { |
||
| 629 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 630 | (__v16hi)_mm256_packus_epi32(__A, __B), |
||
| 631 | (__v16hi)__W); |
||
| 632 | } |
||
| 633 | |||
| 634 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 635 | _mm_maskz_packus_epi16(__mmask16 __M, __m128i __A, __m128i __B) |
||
| 636 | { |
||
| 637 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 638 | (__v16qi)_mm_packus_epi16(__A, __B), |
||
| 639 | (__v16qi)_mm_setzero_si128()); |
||
| 640 | } |
||
| 641 | |||
| 642 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 643 | _mm_mask_packus_epi16(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) |
||
| 644 | { |
||
| 645 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 646 | (__v16qi)_mm_packus_epi16(__A, __B), |
||
| 647 | (__v16qi)__W); |
||
| 648 | } |
||
| 649 | |||
| 650 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 651 | _mm256_maskz_packus_epi16(__mmask32 __M, __m256i __A, __m256i __B) |
||
| 652 | { |
||
| 653 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 654 | (__v32qi)_mm256_packus_epi16(__A, __B), |
||
| 655 | (__v32qi)_mm256_setzero_si256()); |
||
| 656 | } |
||
| 657 | |||
| 658 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 659 | _mm256_mask_packus_epi16(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) |
||
| 660 | { |
||
| 661 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 662 | (__v32qi)_mm256_packus_epi16(__A, __B), |
||
| 663 | (__v32qi)__W); |
||
| 664 | } |
||
| 665 | |||
| 666 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 667 | _mm_mask_adds_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) |
||
| 668 | { |
||
| 669 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 670 | (__v16qi)_mm_adds_epi8(__A, __B), |
||
| 671 | (__v16qi)__W); |
||
| 672 | } |
||
| 673 | |||
| 674 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 675 | _mm_maskz_adds_epi8(__mmask16 __U, __m128i __A, __m128i __B) |
||
| 676 | { |
||
| 677 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 678 | (__v16qi)_mm_adds_epi8(__A, __B), |
||
| 679 | (__v16qi)_mm_setzero_si128()); |
||
| 680 | } |
||
| 681 | |||
| 682 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 683 | _mm256_mask_adds_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) |
||
| 684 | { |
||
| 685 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 686 | (__v32qi)_mm256_adds_epi8(__A, __B), |
||
| 687 | (__v32qi)__W); |
||
| 688 | } |
||
| 689 | |||
| 690 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 691 | _mm256_maskz_adds_epi8(__mmask32 __U, __m256i __A, __m256i __B) |
||
| 692 | { |
||
| 693 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 694 | (__v32qi)_mm256_adds_epi8(__A, __B), |
||
| 695 | (__v32qi)_mm256_setzero_si256()); |
||
| 696 | } |
||
| 697 | |||
| 698 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 699 | _mm_mask_adds_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 700 | { |
||
| 701 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 702 | (__v8hi)_mm_adds_epi16(__A, __B), |
||
| 703 | (__v8hi)__W); |
||
| 704 | } |
||
| 705 | |||
| 706 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 707 | _mm_maskz_adds_epi16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 708 | { |
||
| 709 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 710 | (__v8hi)_mm_adds_epi16(__A, __B), |
||
| 711 | (__v8hi)_mm_setzero_si128()); |
||
| 712 | } |
||
| 713 | |||
| 714 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 715 | _mm256_mask_adds_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 716 | { |
||
| 717 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 718 | (__v16hi)_mm256_adds_epi16(__A, __B), |
||
| 719 | (__v16hi)__W); |
||
| 720 | } |
||
| 721 | |||
| 722 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 723 | _mm256_maskz_adds_epi16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 724 | { |
||
| 725 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 726 | (__v16hi)_mm256_adds_epi16(__A, __B), |
||
| 727 | (__v16hi)_mm256_setzero_si256()); |
||
| 728 | } |
||
| 729 | |||
| 730 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 731 | _mm_mask_adds_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) |
||
| 732 | { |
||
| 733 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 734 | (__v16qi)_mm_adds_epu8(__A, __B), |
||
| 735 | (__v16qi)__W); |
||
| 736 | } |
||
| 737 | |||
| 738 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 739 | _mm_maskz_adds_epu8(__mmask16 __U, __m128i __A, __m128i __B) |
||
| 740 | { |
||
| 741 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 742 | (__v16qi)_mm_adds_epu8(__A, __B), |
||
| 743 | (__v16qi)_mm_setzero_si128()); |
||
| 744 | } |
||
| 745 | |||
| 746 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 747 | _mm256_mask_adds_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) |
||
| 748 | { |
||
| 749 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 750 | (__v32qi)_mm256_adds_epu8(__A, __B), |
||
| 751 | (__v32qi)__W); |
||
| 752 | } |
||
| 753 | |||
| 754 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 755 | _mm256_maskz_adds_epu8(__mmask32 __U, __m256i __A, __m256i __B) |
||
| 756 | { |
||
| 757 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 758 | (__v32qi)_mm256_adds_epu8(__A, __B), |
||
| 759 | (__v32qi)_mm256_setzero_si256()); |
||
| 760 | } |
||
| 761 | |||
| 762 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 763 | _mm_mask_adds_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 764 | { |
||
| 765 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 766 | (__v8hi)_mm_adds_epu16(__A, __B), |
||
| 767 | (__v8hi)__W); |
||
| 768 | } |
||
| 769 | |||
| 770 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 771 | _mm_maskz_adds_epu16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 772 | { |
||
| 773 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 774 | (__v8hi)_mm_adds_epu16(__A, __B), |
||
| 775 | (__v8hi)_mm_setzero_si128()); |
||
| 776 | } |
||
| 777 | |||
| 778 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 779 | _mm256_mask_adds_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 780 | { |
||
| 781 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 782 | (__v16hi)_mm256_adds_epu16(__A, __B), |
||
| 783 | (__v16hi)__W); |
||
| 784 | } |
||
| 785 | |||
| 786 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 787 | _mm256_maskz_adds_epu16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 788 | { |
||
| 789 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 790 | (__v16hi)_mm256_adds_epu16(__A, __B), |
||
| 791 | (__v16hi)_mm256_setzero_si256()); |
||
| 792 | } |
||
| 793 | |||
| 794 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 795 | _mm_mask_avg_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) |
||
| 796 | { |
||
| 797 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 798 | (__v16qi)_mm_avg_epu8(__A, __B), |
||
| 799 | (__v16qi)__W); |
||
| 800 | } |
||
| 801 | |||
| 802 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 803 | _mm_maskz_avg_epu8(__mmask16 __U, __m128i __A, __m128i __B) |
||
| 804 | { |
||
| 805 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 806 | (__v16qi)_mm_avg_epu8(__A, __B), |
||
| 807 | (__v16qi)_mm_setzero_si128()); |
||
| 808 | } |
||
| 809 | |||
| 810 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 811 | _mm256_mask_avg_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) |
||
| 812 | { |
||
| 813 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 814 | (__v32qi)_mm256_avg_epu8(__A, __B), |
||
| 815 | (__v32qi)__W); |
||
| 816 | } |
||
| 817 | |||
| 818 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 819 | _mm256_maskz_avg_epu8(__mmask32 __U, __m256i __A, __m256i __B) |
||
| 820 | { |
||
| 821 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 822 | (__v32qi)_mm256_avg_epu8(__A, __B), |
||
| 823 | (__v32qi)_mm256_setzero_si256()); |
||
| 824 | } |
||
| 825 | |||
| 826 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 827 | _mm_mask_avg_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 828 | { |
||
| 829 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 830 | (__v8hi)_mm_avg_epu16(__A, __B), |
||
| 831 | (__v8hi)__W); |
||
| 832 | } |
||
| 833 | |||
| 834 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 835 | _mm_maskz_avg_epu16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 836 | { |
||
| 837 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 838 | (__v8hi)_mm_avg_epu16(__A, __B), |
||
| 839 | (__v8hi)_mm_setzero_si128()); |
||
| 840 | } |
||
| 841 | |||
| 842 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 843 | _mm256_mask_avg_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 844 | { |
||
| 845 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 846 | (__v16hi)_mm256_avg_epu16(__A, __B), |
||
| 847 | (__v16hi)__W); |
||
| 848 | } |
||
| 849 | |||
| 850 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 851 | _mm256_maskz_avg_epu16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 852 | { |
||
| 853 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 854 | (__v16hi)_mm256_avg_epu16(__A, __B), |
||
| 855 | (__v16hi)_mm256_setzero_si256()); |
||
| 856 | } |
||
| 857 | |||
| 858 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 859 | _mm_maskz_max_epi8(__mmask16 __M, __m128i __A, __m128i __B) |
||
| 860 | { |
||
| 861 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 862 | (__v16qi)_mm_max_epi8(__A, __B), |
||
| 863 | (__v16qi)_mm_setzero_si128()); |
||
| 864 | } |
||
| 865 | |||
| 866 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 867 | _mm_mask_max_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) |
||
| 868 | { |
||
| 869 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 870 | (__v16qi)_mm_max_epi8(__A, __B), |
||
| 871 | (__v16qi)__W); |
||
| 872 | } |
||
| 873 | |||
| 874 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 875 | _mm256_maskz_max_epi8(__mmask32 __M, __m256i __A, __m256i __B) |
||
| 876 | { |
||
| 877 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 878 | (__v32qi)_mm256_max_epi8(__A, __B), |
||
| 879 | (__v32qi)_mm256_setzero_si256()); |
||
| 880 | } |
||
| 881 | |||
| 882 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 883 | _mm256_mask_max_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) |
||
| 884 | { |
||
| 885 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 886 | (__v32qi)_mm256_max_epi8(__A, __B), |
||
| 887 | (__v32qi)__W); |
||
| 888 | } |
||
| 889 | |||
| 890 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 891 | _mm_maskz_max_epi16(__mmask8 __M, __m128i __A, __m128i __B) |
||
| 892 | { |
||
| 893 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 894 | (__v8hi)_mm_max_epi16(__A, __B), |
||
| 895 | (__v8hi)_mm_setzero_si128()); |
||
| 896 | } |
||
| 897 | |||
| 898 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 899 | _mm_mask_max_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
||
| 900 | { |
||
| 901 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 902 | (__v8hi)_mm_max_epi16(__A, __B), |
||
| 903 | (__v8hi)__W); |
||
| 904 | } |
||
| 905 | |||
| 906 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 907 | _mm256_maskz_max_epi16(__mmask16 __M, __m256i __A, __m256i __B) |
||
| 908 | { |
||
| 909 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 910 | (__v16hi)_mm256_max_epi16(__A, __B), |
||
| 911 | (__v16hi)_mm256_setzero_si256()); |
||
| 912 | } |
||
| 913 | |||
| 914 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 915 | _mm256_mask_max_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) |
||
| 916 | { |
||
| 917 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 918 | (__v16hi)_mm256_max_epi16(__A, __B), |
||
| 919 | (__v16hi)__W); |
||
| 920 | } |
||
| 921 | |||
| 922 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 923 | _mm_maskz_max_epu8(__mmask16 __M, __m128i __A, __m128i __B) |
||
| 924 | { |
||
| 925 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 926 | (__v16qi)_mm_max_epu8(__A, __B), |
||
| 927 | (__v16qi)_mm_setzero_si128()); |
||
| 928 | } |
||
| 929 | |||
| 930 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 931 | _mm_mask_max_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) |
||
| 932 | { |
||
| 933 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 934 | (__v16qi)_mm_max_epu8(__A, __B), |
||
| 935 | (__v16qi)__W); |
||
| 936 | } |
||
| 937 | |||
| 938 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 939 | _mm256_maskz_max_epu8 (__mmask32 __M, __m256i __A, __m256i __B) |
||
| 940 | { |
||
| 941 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 942 | (__v32qi)_mm256_max_epu8(__A, __B), |
||
| 943 | (__v32qi)_mm256_setzero_si256()); |
||
| 944 | } |
||
| 945 | |||
| 946 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 947 | _mm256_mask_max_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) |
||
| 948 | { |
||
| 949 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 950 | (__v32qi)_mm256_max_epu8(__A, __B), |
||
| 951 | (__v32qi)__W); |
||
| 952 | } |
||
| 953 | |||
| 954 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 955 | _mm_maskz_max_epu16(__mmask8 __M, __m128i __A, __m128i __B) |
||
| 956 | { |
||
| 957 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 958 | (__v8hi)_mm_max_epu16(__A, __B), |
||
| 959 | (__v8hi)_mm_setzero_si128()); |
||
| 960 | } |
||
| 961 | |||
| 962 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 963 | _mm_mask_max_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
||
| 964 | { |
||
| 965 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 966 | (__v8hi)_mm_max_epu16(__A, __B), |
||
| 967 | (__v8hi)__W); |
||
| 968 | } |
||
| 969 | |||
| 970 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 971 | _mm256_maskz_max_epu16(__mmask16 __M, __m256i __A, __m256i __B) |
||
| 972 | { |
||
| 973 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 974 | (__v16hi)_mm256_max_epu16(__A, __B), |
||
| 975 | (__v16hi)_mm256_setzero_si256()); |
||
| 976 | } |
||
| 977 | |||
| 978 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 979 | _mm256_mask_max_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) |
||
| 980 | { |
||
| 981 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 982 | (__v16hi)_mm256_max_epu16(__A, __B), |
||
| 983 | (__v16hi)__W); |
||
| 984 | } |
||
| 985 | |||
| 986 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 987 | _mm_maskz_min_epi8(__mmask16 __M, __m128i __A, __m128i __B) |
||
| 988 | { |
||
| 989 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 990 | (__v16qi)_mm_min_epi8(__A, __B), |
||
| 991 | (__v16qi)_mm_setzero_si128()); |
||
| 992 | } |
||
| 993 | |||
| 994 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 995 | _mm_mask_min_epi8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) |
||
| 996 | { |
||
| 997 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 998 | (__v16qi)_mm_min_epi8(__A, __B), |
||
| 999 | (__v16qi)__W); |
||
| 1000 | } |
||
| 1001 | |||
| 1002 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1003 | _mm256_maskz_min_epi8(__mmask32 __M, __m256i __A, __m256i __B) |
||
| 1004 | { |
||
| 1005 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 1006 | (__v32qi)_mm256_min_epi8(__A, __B), |
||
| 1007 | (__v32qi)_mm256_setzero_si256()); |
||
| 1008 | } |
||
| 1009 | |||
| 1010 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1011 | _mm256_mask_min_epi8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) |
||
| 1012 | { |
||
| 1013 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 1014 | (__v32qi)_mm256_min_epi8(__A, __B), |
||
| 1015 | (__v32qi)__W); |
||
| 1016 | } |
||
| 1017 | |||
| 1018 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1019 | _mm_maskz_min_epi16(__mmask8 __M, __m128i __A, __m128i __B) |
||
| 1020 | { |
||
| 1021 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 1022 | (__v8hi)_mm_min_epi16(__A, __B), |
||
| 1023 | (__v8hi)_mm_setzero_si128()); |
||
| 1024 | } |
||
| 1025 | |||
| 1026 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1027 | _mm_mask_min_epi16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
||
| 1028 | { |
||
| 1029 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 1030 | (__v8hi)_mm_min_epi16(__A, __B), |
||
| 1031 | (__v8hi)__W); |
||
| 1032 | } |
||
| 1033 | |||
| 1034 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1035 | _mm256_maskz_min_epi16(__mmask16 __M, __m256i __A, __m256i __B) |
||
| 1036 | { |
||
| 1037 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 1038 | (__v16hi)_mm256_min_epi16(__A, __B), |
||
| 1039 | (__v16hi)_mm256_setzero_si256()); |
||
| 1040 | } |
||
| 1041 | |||
| 1042 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1043 | _mm256_mask_min_epi16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) |
||
| 1044 | { |
||
| 1045 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 1046 | (__v16hi)_mm256_min_epi16(__A, __B), |
||
| 1047 | (__v16hi)__W); |
||
| 1048 | } |
||
| 1049 | |||
| 1050 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1051 | _mm_maskz_min_epu8(__mmask16 __M, __m128i __A, __m128i __B) |
||
| 1052 | { |
||
| 1053 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 1054 | (__v16qi)_mm_min_epu8(__A, __B), |
||
| 1055 | (__v16qi)_mm_setzero_si128()); |
||
| 1056 | } |
||
| 1057 | |||
| 1058 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1059 | _mm_mask_min_epu8(__m128i __W, __mmask16 __M, __m128i __A, __m128i __B) |
||
| 1060 | { |
||
| 1061 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 1062 | (__v16qi)_mm_min_epu8(__A, __B), |
||
| 1063 | (__v16qi)__W); |
||
| 1064 | } |
||
| 1065 | |||
| 1066 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1067 | _mm256_maskz_min_epu8 (__mmask32 __M, __m256i __A, __m256i __B) |
||
| 1068 | { |
||
| 1069 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 1070 | (__v32qi)_mm256_min_epu8(__A, __B), |
||
| 1071 | (__v32qi)_mm256_setzero_si256()); |
||
| 1072 | } |
||
| 1073 | |||
| 1074 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1075 | _mm256_mask_min_epu8(__m256i __W, __mmask32 __M, __m256i __A, __m256i __B) |
||
| 1076 | { |
||
| 1077 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__M, |
||
| 1078 | (__v32qi)_mm256_min_epu8(__A, __B), |
||
| 1079 | (__v32qi)__W); |
||
| 1080 | } |
||
| 1081 | |||
| 1082 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1083 | _mm_maskz_min_epu16(__mmask8 __M, __m128i __A, __m128i __B) |
||
| 1084 | { |
||
| 1085 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 1086 | (__v8hi)_mm_min_epu16(__A, __B), |
||
| 1087 | (__v8hi)_mm_setzero_si128()); |
||
| 1088 | } |
||
| 1089 | |||
| 1090 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1091 | _mm_mask_min_epu16(__m128i __W, __mmask8 __M, __m128i __A, __m128i __B) |
||
| 1092 | { |
||
| 1093 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 1094 | (__v8hi)_mm_min_epu16(__A, __B), |
||
| 1095 | (__v8hi)__W); |
||
| 1096 | } |
||
| 1097 | |||
| 1098 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1099 | _mm256_maskz_min_epu16(__mmask16 __M, __m256i __A, __m256i __B) |
||
| 1100 | { |
||
| 1101 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 1102 | (__v16hi)_mm256_min_epu16(__A, __B), |
||
| 1103 | (__v16hi)_mm256_setzero_si256()); |
||
| 1104 | } |
||
| 1105 | |||
| 1106 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1107 | _mm256_mask_min_epu16(__m256i __W, __mmask16 __M, __m256i __A, __m256i __B) |
||
| 1108 | { |
||
| 1109 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 1110 | (__v16hi)_mm256_min_epu16(__A, __B), |
||
| 1111 | (__v16hi)__W); |
||
| 1112 | } |
||
| 1113 | |||
| 1114 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1115 | _mm_mask_shuffle_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) |
||
| 1116 | { |
||
| 1117 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1118 | (__v16qi)_mm_shuffle_epi8(__A, __B), |
||
| 1119 | (__v16qi)__W); |
||
| 1120 | } |
||
| 1121 | |||
| 1122 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1123 | _mm_maskz_shuffle_epi8(__mmask16 __U, __m128i __A, __m128i __B) |
||
| 1124 | { |
||
| 1125 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1126 | (__v16qi)_mm_shuffle_epi8(__A, __B), |
||
| 1127 | (__v16qi)_mm_setzero_si128()); |
||
| 1128 | } |
||
| 1129 | |||
| 1130 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1131 | _mm256_mask_shuffle_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) |
||
| 1132 | { |
||
| 1133 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1134 | (__v32qi)_mm256_shuffle_epi8(__A, __B), |
||
| 1135 | (__v32qi)__W); |
||
| 1136 | } |
||
| 1137 | |||
| 1138 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1139 | _mm256_maskz_shuffle_epi8(__mmask32 __U, __m256i __A, __m256i __B) |
||
| 1140 | { |
||
| 1141 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1142 | (__v32qi)_mm256_shuffle_epi8(__A, __B), |
||
| 1143 | (__v32qi)_mm256_setzero_si256()); |
||
| 1144 | } |
||
| 1145 | |||
| 1146 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1147 | _mm_mask_subs_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) |
||
| 1148 | { |
||
| 1149 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1150 | (__v16qi)_mm_subs_epi8(__A, __B), |
||
| 1151 | (__v16qi)__W); |
||
| 1152 | } |
||
| 1153 | |||
| 1154 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1155 | _mm_maskz_subs_epi8(__mmask16 __U, __m128i __A, __m128i __B) |
||
| 1156 | { |
||
| 1157 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1158 | (__v16qi)_mm_subs_epi8(__A, __B), |
||
| 1159 | (__v16qi)_mm_setzero_si128()); |
||
| 1160 | } |
||
| 1161 | |||
| 1162 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1163 | _mm256_mask_subs_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) |
||
| 1164 | { |
||
| 1165 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1166 | (__v32qi)_mm256_subs_epi8(__A, __B), |
||
| 1167 | (__v32qi)__W); |
||
| 1168 | } |
||
| 1169 | |||
| 1170 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1171 | _mm256_maskz_subs_epi8(__mmask32 __U, __m256i __A, __m256i __B) |
||
| 1172 | { |
||
| 1173 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1174 | (__v32qi)_mm256_subs_epi8(__A, __B), |
||
| 1175 | (__v32qi)_mm256_setzero_si256()); |
||
| 1176 | } |
||
| 1177 | |||
| 1178 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1179 | _mm_mask_subs_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 1180 | { |
||
| 1181 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1182 | (__v8hi)_mm_subs_epi16(__A, __B), |
||
| 1183 | (__v8hi)__W); |
||
| 1184 | } |
||
| 1185 | |||
| 1186 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1187 | _mm_maskz_subs_epi16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 1188 | { |
||
| 1189 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1190 | (__v8hi)_mm_subs_epi16(__A, __B), |
||
| 1191 | (__v8hi)_mm_setzero_si128()); |
||
| 1192 | } |
||
| 1193 | |||
| 1194 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1195 | _mm256_mask_subs_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 1196 | { |
||
| 1197 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1198 | (__v16hi)_mm256_subs_epi16(__A, __B), |
||
| 1199 | (__v16hi)__W); |
||
| 1200 | } |
||
| 1201 | |||
| 1202 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1203 | _mm256_maskz_subs_epi16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 1204 | { |
||
| 1205 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1206 | (__v16hi)_mm256_subs_epi16(__A, __B), |
||
| 1207 | (__v16hi)_mm256_setzero_si256()); |
||
| 1208 | } |
||
| 1209 | |||
| 1210 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1211 | _mm_mask_subs_epu8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) |
||
| 1212 | { |
||
| 1213 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1214 | (__v16qi)_mm_subs_epu8(__A, __B), |
||
| 1215 | (__v16qi)__W); |
||
| 1216 | } |
||
| 1217 | |||
| 1218 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1219 | _mm_maskz_subs_epu8(__mmask16 __U, __m128i __A, __m128i __B) |
||
| 1220 | { |
||
| 1221 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1222 | (__v16qi)_mm_subs_epu8(__A, __B), |
||
| 1223 | (__v16qi)_mm_setzero_si128()); |
||
| 1224 | } |
||
| 1225 | |||
| 1226 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1227 | _mm256_mask_subs_epu8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) |
||
| 1228 | { |
||
| 1229 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1230 | (__v32qi)_mm256_subs_epu8(__A, __B), |
||
| 1231 | (__v32qi)__W); |
||
| 1232 | } |
||
| 1233 | |||
| 1234 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1235 | _mm256_maskz_subs_epu8(__mmask32 __U, __m256i __A, __m256i __B) |
||
| 1236 | { |
||
| 1237 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1238 | (__v32qi)_mm256_subs_epu8(__A, __B), |
||
| 1239 | (__v32qi)_mm256_setzero_si256()); |
||
| 1240 | } |
||
| 1241 | |||
| 1242 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1243 | _mm_mask_subs_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 1244 | { |
||
| 1245 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1246 | (__v8hi)_mm_subs_epu16(__A, __B), |
||
| 1247 | (__v8hi)__W); |
||
| 1248 | } |
||
| 1249 | |||
| 1250 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1251 | _mm_maskz_subs_epu16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 1252 | { |
||
| 1253 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1254 | (__v8hi)_mm_subs_epu16(__A, __B), |
||
| 1255 | (__v8hi)_mm_setzero_si128()); |
||
| 1256 | } |
||
| 1257 | |||
| 1258 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1259 | _mm256_mask_subs_epu16(__m256i __W, __mmask16 __U, __m256i __A, |
||
| 1260 | __m256i __B) { |
||
| 1261 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1262 | (__v16hi)_mm256_subs_epu16(__A, __B), |
||
| 1263 | (__v16hi)__W); |
||
| 1264 | } |
||
| 1265 | |||
| 1266 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1267 | _mm256_maskz_subs_epu16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 1268 | { |
||
| 1269 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1270 | (__v16hi)_mm256_subs_epu16(__A, __B), |
||
| 1271 | (__v16hi)_mm256_setzero_si256()); |
||
| 1272 | } |
||
| 1273 | |||
| 1274 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1275 | _mm_permutex2var_epi16(__m128i __A, __m128i __I, __m128i __B) |
||
| 1276 | { |
||
| 1277 | return (__m128i)__builtin_ia32_vpermi2varhi128((__v8hi)__A, (__v8hi)__I, |
||
| 1278 | (__v8hi) __B); |
||
| 1279 | } |
||
| 1280 | |||
| 1281 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1282 | _mm_mask_permutex2var_epi16(__m128i __A, __mmask8 __U, __m128i __I, |
||
| 1283 | __m128i __B) |
||
| 1284 | { |
||
| 1285 | return (__m128i)__builtin_ia32_selectw_128(__U, |
||
| 1286 | (__v8hi)_mm_permutex2var_epi16(__A, __I, __B), |
||
| 1287 | (__v8hi)__A); |
||
| 1288 | } |
||
| 1289 | |||
| 1290 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1291 | _mm_mask2_permutex2var_epi16(__m128i __A, __m128i __I, __mmask8 __U, |
||
| 1292 | __m128i __B) |
||
| 1293 | { |
||
| 1294 | return (__m128i)__builtin_ia32_selectw_128(__U, |
||
| 1295 | (__v8hi)_mm_permutex2var_epi16(__A, __I, __B), |
||
| 1296 | (__v8hi)__I); |
||
| 1297 | } |
||
| 1298 | |||
| 1299 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1300 | _mm_maskz_permutex2var_epi16 (__mmask8 __U, __m128i __A, __m128i __I, |
||
| 1301 | __m128i __B) |
||
| 1302 | { |
||
| 1303 | return (__m128i)__builtin_ia32_selectw_128(__U, |
||
| 1304 | (__v8hi)_mm_permutex2var_epi16(__A, __I, __B), |
||
| 1305 | (__v8hi)_mm_setzero_si128()); |
||
| 1306 | } |
||
| 1307 | |||
| 1308 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1309 | _mm256_permutex2var_epi16(__m256i __A, __m256i __I, __m256i __B) |
||
| 1310 | { |
||
| 1311 | return (__m256i)__builtin_ia32_vpermi2varhi256((__v16hi)__A, (__v16hi)__I, |
||
| 1312 | (__v16hi)__B); |
||
| 1313 | } |
||
| 1314 | |||
| 1315 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1316 | _mm256_mask_permutex2var_epi16(__m256i __A, __mmask16 __U, __m256i __I, |
||
| 1317 | __m256i __B) |
||
| 1318 | { |
||
| 1319 | return (__m256i)__builtin_ia32_selectw_256(__U, |
||
| 1320 | (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B), |
||
| 1321 | (__v16hi)__A); |
||
| 1322 | } |
||
| 1323 | |||
| 1324 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1325 | _mm256_mask2_permutex2var_epi16(__m256i __A, __m256i __I, __mmask16 __U, |
||
| 1326 | __m256i __B) |
||
| 1327 | { |
||
| 1328 | return (__m256i)__builtin_ia32_selectw_256(__U, |
||
| 1329 | (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B), |
||
| 1330 | (__v16hi)__I); |
||
| 1331 | } |
||
| 1332 | |||
| 1333 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1334 | _mm256_maskz_permutex2var_epi16 (__mmask16 __U, __m256i __A, __m256i __I, |
||
| 1335 | __m256i __B) |
||
| 1336 | { |
||
| 1337 | return (__m256i)__builtin_ia32_selectw_256(__U, |
||
| 1338 | (__v16hi)_mm256_permutex2var_epi16(__A, __I, __B), |
||
| 1339 | (__v16hi)_mm256_setzero_si256()); |
||
| 1340 | } |
||
| 1341 | |||
| 1342 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1343 | _mm_mask_maddubs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) { |
||
| 1344 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1345 | (__v8hi)_mm_maddubs_epi16(__X, __Y), |
||
| 1346 | (__v8hi)__W); |
||
| 1347 | } |
||
| 1348 | |||
| 1349 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1350 | _mm_maskz_maddubs_epi16(__mmask8 __U, __m128i __X, __m128i __Y) { |
||
| 1351 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1352 | (__v8hi)_mm_maddubs_epi16(__X, __Y), |
||
| 1353 | (__v8hi)_mm_setzero_si128()); |
||
| 1354 | } |
||
| 1355 | |||
| 1356 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1357 | _mm256_mask_maddubs_epi16(__m256i __W, __mmask16 __U, __m256i __X, |
||
| 1358 | __m256i __Y) { |
||
| 1359 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1360 | (__v16hi)_mm256_maddubs_epi16(__X, __Y), |
||
| 1361 | (__v16hi)__W); |
||
| 1362 | } |
||
| 1363 | |||
| 1364 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1365 | _mm256_maskz_maddubs_epi16(__mmask16 __U, __m256i __X, __m256i __Y) { |
||
| 1366 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1367 | (__v16hi)_mm256_maddubs_epi16(__X, __Y), |
||
| 1368 | (__v16hi)_mm256_setzero_si256()); |
||
| 1369 | } |
||
| 1370 | |||
| 1371 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1372 | _mm_mask_madd_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1373 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
||
| 1374 | (__v4si)_mm_madd_epi16(__A, __B), |
||
| 1375 | (__v4si)__W); |
||
| 1376 | } |
||
| 1377 | |||
| 1378 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1379 | _mm_maskz_madd_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1380 | return (__m128i)__builtin_ia32_selectd_128((__mmask8)__U, |
||
| 1381 | (__v4si)_mm_madd_epi16(__A, __B), |
||
| 1382 | (__v4si)_mm_setzero_si128()); |
||
| 1383 | } |
||
| 1384 | |||
| 1385 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1386 | _mm256_mask_madd_epi16(__m256i __W, __mmask8 __U, __m256i __A, __m256i __B) { |
||
| 1387 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
||
| 1388 | (__v8si)_mm256_madd_epi16(__A, __B), |
||
| 1389 | (__v8si)__W); |
||
| 1390 | } |
||
| 1391 | |||
| 1392 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1393 | _mm256_maskz_madd_epi16(__mmask8 __U, __m256i __A, __m256i __B) { |
||
| 1394 | return (__m256i)__builtin_ia32_selectd_256((__mmask8)__U, |
||
| 1395 | (__v8si)_mm256_madd_epi16(__A, __B), |
||
| 1396 | (__v8si)_mm256_setzero_si256()); |
||
| 1397 | } |
||
| 1398 | |||
| 1399 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1400 | _mm_cvtsepi16_epi8 (__m128i __A) { |
||
| 1401 | return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A, |
||
| 1402 | (__v16qi) _mm_setzero_si128(), |
||
| 1403 | (__mmask8) -1); |
||
| 1404 | } |
||
| 1405 | |||
| 1406 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1407 | _mm_mask_cvtsepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) { |
||
| 1408 | return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A, |
||
| 1409 | (__v16qi) __O, |
||
| 1410 | __M); |
||
| 1411 | } |
||
| 1412 | |||
| 1413 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1414 | _mm_maskz_cvtsepi16_epi8 (__mmask8 __M, __m128i __A) { |
||
| 1415 | return (__m128i) __builtin_ia32_pmovswb128_mask ((__v8hi) __A, |
||
| 1416 | (__v16qi) _mm_setzero_si128(), |
||
| 1417 | __M); |
||
| 1418 | } |
||
| 1419 | |||
| 1420 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1421 | _mm256_cvtsepi16_epi8 (__m256i __A) { |
||
| 1422 | return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A, |
||
| 1423 | (__v16qi) _mm_setzero_si128(), |
||
| 1424 | (__mmask16) -1); |
||
| 1425 | } |
||
| 1426 | |||
| 1427 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1428 | _mm256_mask_cvtsepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) { |
||
| 1429 | return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A, |
||
| 1430 | (__v16qi) __O, |
||
| 1431 | __M); |
||
| 1432 | } |
||
| 1433 | |||
| 1434 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1435 | _mm256_maskz_cvtsepi16_epi8 (__mmask16 __M, __m256i __A) { |
||
| 1436 | return (__m128i) __builtin_ia32_pmovswb256_mask ((__v16hi) __A, |
||
| 1437 | (__v16qi) _mm_setzero_si128(), |
||
| 1438 | __M); |
||
| 1439 | } |
||
| 1440 | |||
| 1441 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1442 | _mm_cvtusepi16_epi8 (__m128i __A) { |
||
| 1443 | return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A, |
||
| 1444 | (__v16qi) _mm_setzero_si128(), |
||
| 1445 | (__mmask8) -1); |
||
| 1446 | } |
||
| 1447 | |||
| 1448 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1449 | _mm_mask_cvtusepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) { |
||
| 1450 | return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A, |
||
| 1451 | (__v16qi) __O, |
||
| 1452 | __M); |
||
| 1453 | } |
||
| 1454 | |||
| 1455 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1456 | _mm_maskz_cvtusepi16_epi8 (__mmask8 __M, __m128i __A) { |
||
| 1457 | return (__m128i) __builtin_ia32_pmovuswb128_mask ((__v8hi) __A, |
||
| 1458 | (__v16qi) _mm_setzero_si128(), |
||
| 1459 | __M); |
||
| 1460 | } |
||
| 1461 | |||
| 1462 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1463 | _mm256_cvtusepi16_epi8 (__m256i __A) { |
||
| 1464 | return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A, |
||
| 1465 | (__v16qi) _mm_setzero_si128(), |
||
| 1466 | (__mmask16) -1); |
||
| 1467 | } |
||
| 1468 | |||
| 1469 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1470 | _mm256_mask_cvtusepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) { |
||
| 1471 | return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A, |
||
| 1472 | (__v16qi) __O, |
||
| 1473 | __M); |
||
| 1474 | } |
||
| 1475 | |||
| 1476 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1477 | _mm256_maskz_cvtusepi16_epi8 (__mmask16 __M, __m256i __A) { |
||
| 1478 | return (__m128i) __builtin_ia32_pmovuswb256_mask ((__v16hi) __A, |
||
| 1479 | (__v16qi) _mm_setzero_si128(), |
||
| 1480 | __M); |
||
| 1481 | } |
||
| 1482 | |||
| 1483 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1484 | _mm_cvtepi16_epi8 (__m128i __A) { |
||
| 1485 | return (__m128i)__builtin_shufflevector( |
||
| 1486 | __builtin_convertvector((__v8hi)__A, __v8qi), |
||
| 1487 | (__v8qi){0, 0, 0, 0, 0, 0, 0, 0}, 0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, |
||
| 1488 | 12, 13, 14, 15); |
||
| 1489 | } |
||
| 1490 | |||
| 1491 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1492 | _mm_mask_cvtepi16_epi8 (__m128i __O, __mmask8 __M, __m128i __A) { |
||
| 1493 | return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A, |
||
| 1494 | (__v16qi) __O, |
||
| 1495 | __M); |
||
| 1496 | } |
||
| 1497 | |||
| 1498 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1499 | _mm_maskz_cvtepi16_epi8 (__mmask8 __M, __m128i __A) { |
||
| 1500 | return (__m128i) __builtin_ia32_pmovwb128_mask ((__v8hi) __A, |
||
| 1501 | (__v16qi) _mm_setzero_si128(), |
||
| 1502 | __M); |
||
| 1503 | } |
||
| 1504 | |||
| 1505 | static __inline__ void __DEFAULT_FN_ATTRS128 |
||
| 1506 | _mm_mask_cvtepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
||
| 1507 | { |
||
| 1508 | __builtin_ia32_pmovwb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M); |
||
| 1509 | } |
||
| 1510 | |||
| 1511 | |||
| 1512 | static __inline__ void __DEFAULT_FN_ATTRS128 |
||
| 1513 | _mm_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
||
| 1514 | { |
||
| 1515 | __builtin_ia32_pmovswb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M); |
||
| 1516 | } |
||
| 1517 | |||
| 1518 | static __inline__ void __DEFAULT_FN_ATTRS128 |
||
| 1519 | _mm_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask8 __M, __m128i __A) |
||
| 1520 | { |
||
| 1521 | __builtin_ia32_pmovuswb128mem_mask ((__v16qi *) __P, (__v8hi) __A, __M); |
||
| 1522 | } |
||
| 1523 | |||
| 1524 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1525 | _mm256_cvtepi16_epi8 (__m256i __A) { |
||
| 1526 | return (__m128i)__builtin_convertvector((__v16hi) __A, __v16qi); |
||
| 1527 | } |
||
| 1528 | |||
| 1529 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1530 | _mm256_mask_cvtepi16_epi8 (__m128i __O, __mmask16 __M, __m256i __A) { |
||
| 1531 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 1532 | (__v16qi)_mm256_cvtepi16_epi8(__A), |
||
| 1533 | (__v16qi)__O); |
||
| 1534 | } |
||
| 1535 | |||
| 1536 | static __inline__ __m128i __DEFAULT_FN_ATTRS256 |
||
| 1537 | _mm256_maskz_cvtepi16_epi8 (__mmask16 __M, __m256i __A) { |
||
| 1538 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__M, |
||
| 1539 | (__v16qi)_mm256_cvtepi16_epi8(__A), |
||
| 1540 | (__v16qi)_mm_setzero_si128()); |
||
| 1541 | } |
||
| 1542 | |||
| 1543 | static __inline__ void __DEFAULT_FN_ATTRS256 |
||
| 1544 | _mm256_mask_cvtepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A) |
||
| 1545 | { |
||
| 1546 | __builtin_ia32_pmovwb256mem_mask ((__v16qi *) __P, (__v16hi) __A, __M); |
||
| 1547 | } |
||
| 1548 | |||
| 1549 | static __inline__ void __DEFAULT_FN_ATTRS256 |
||
| 1550 | _mm256_mask_cvtsepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A) |
||
| 1551 | { |
||
| 1552 | __builtin_ia32_pmovswb256mem_mask ((__v16qi *) __P, (__v16hi) __A, __M); |
||
| 1553 | } |
||
| 1554 | |||
| 1555 | static __inline__ void __DEFAULT_FN_ATTRS256 |
||
| 1556 | _mm256_mask_cvtusepi16_storeu_epi8 (void * __P, __mmask16 __M, __m256i __A) |
||
| 1557 | { |
||
| 1558 | __builtin_ia32_pmovuswb256mem_mask ((__v16qi*) __P, (__v16hi) __A, __M); |
||
| 1559 | } |
||
| 1560 | |||
| 1561 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1562 | _mm_mask_mulhrs_epi16(__m128i __W, __mmask8 __U, __m128i __X, __m128i __Y) { |
||
| 1563 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1564 | (__v8hi)_mm_mulhrs_epi16(__X, __Y), |
||
| 1565 | (__v8hi)__W); |
||
| 1566 | } |
||
| 1567 | |||
| 1568 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1569 | _mm_maskz_mulhrs_epi16(__mmask8 __U, __m128i __X, __m128i __Y) { |
||
| 1570 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1571 | (__v8hi)_mm_mulhrs_epi16(__X, __Y), |
||
| 1572 | (__v8hi)_mm_setzero_si128()); |
||
| 1573 | } |
||
| 1574 | |||
| 1575 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1576 | _mm256_mask_mulhrs_epi16(__m256i __W, __mmask16 __U, __m256i __X, __m256i __Y) { |
||
| 1577 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1578 | (__v16hi)_mm256_mulhrs_epi16(__X, __Y), |
||
| 1579 | (__v16hi)__W); |
||
| 1580 | } |
||
| 1581 | |||
| 1582 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1583 | _mm256_maskz_mulhrs_epi16(__mmask16 __U, __m256i __X, __m256i __Y) { |
||
| 1584 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1585 | (__v16hi)_mm256_mulhrs_epi16(__X, __Y), |
||
| 1586 | (__v16hi)_mm256_setzero_si256()); |
||
| 1587 | } |
||
| 1588 | |||
| 1589 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1590 | _mm_mask_mulhi_epu16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1591 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1592 | (__v8hi)_mm_mulhi_epu16(__A, __B), |
||
| 1593 | (__v8hi)__W); |
||
| 1594 | } |
||
| 1595 | |||
| 1596 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1597 | _mm_maskz_mulhi_epu16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1598 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1599 | (__v8hi)_mm_mulhi_epu16(__A, __B), |
||
| 1600 | (__v8hi)_mm_setzero_si128()); |
||
| 1601 | } |
||
| 1602 | |||
| 1603 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1604 | _mm256_mask_mulhi_epu16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1605 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1606 | (__v16hi)_mm256_mulhi_epu16(__A, __B), |
||
| 1607 | (__v16hi)__W); |
||
| 1608 | } |
||
| 1609 | |||
| 1610 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1611 | _mm256_maskz_mulhi_epu16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1612 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1613 | (__v16hi)_mm256_mulhi_epu16(__A, __B), |
||
| 1614 | (__v16hi)_mm256_setzero_si256()); |
||
| 1615 | } |
||
| 1616 | |||
| 1617 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1618 | _mm_mask_mulhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1619 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1620 | (__v8hi)_mm_mulhi_epi16(__A, __B), |
||
| 1621 | (__v8hi)__W); |
||
| 1622 | } |
||
| 1623 | |||
| 1624 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1625 | _mm_maskz_mulhi_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1626 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1627 | (__v8hi)_mm_mulhi_epi16(__A, __B), |
||
| 1628 | (__v8hi)_mm_setzero_si128()); |
||
| 1629 | } |
||
| 1630 | |||
| 1631 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1632 | _mm256_mask_mulhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1633 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1634 | (__v16hi)_mm256_mulhi_epi16(__A, __B), |
||
| 1635 | (__v16hi)__W); |
||
| 1636 | } |
||
| 1637 | |||
| 1638 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1639 | _mm256_maskz_mulhi_epi16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1640 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1641 | (__v16hi)_mm256_mulhi_epi16(__A, __B), |
||
| 1642 | (__v16hi)_mm256_setzero_si256()); |
||
| 1643 | } |
||
| 1644 | |||
| 1645 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1646 | _mm_mask_unpackhi_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) { |
||
| 1647 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1648 | (__v16qi)_mm_unpackhi_epi8(__A, __B), |
||
| 1649 | (__v16qi)__W); |
||
| 1650 | } |
||
| 1651 | |||
| 1652 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1653 | _mm_maskz_unpackhi_epi8(__mmask16 __U, __m128i __A, __m128i __B) { |
||
| 1654 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1655 | (__v16qi)_mm_unpackhi_epi8(__A, __B), |
||
| 1656 | (__v16qi)_mm_setzero_si128()); |
||
| 1657 | } |
||
| 1658 | |||
| 1659 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1660 | _mm256_mask_unpackhi_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) { |
||
| 1661 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1662 | (__v32qi)_mm256_unpackhi_epi8(__A, __B), |
||
| 1663 | (__v32qi)__W); |
||
| 1664 | } |
||
| 1665 | |||
| 1666 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1667 | _mm256_maskz_unpackhi_epi8(__mmask32 __U, __m256i __A, __m256i __B) { |
||
| 1668 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1669 | (__v32qi)_mm256_unpackhi_epi8(__A, __B), |
||
| 1670 | (__v32qi)_mm256_setzero_si256()); |
||
| 1671 | } |
||
| 1672 | |||
| 1673 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1674 | _mm_mask_unpackhi_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1675 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1676 | (__v8hi)_mm_unpackhi_epi16(__A, __B), |
||
| 1677 | (__v8hi)__W); |
||
| 1678 | } |
||
| 1679 | |||
| 1680 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1681 | _mm_maskz_unpackhi_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1682 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1683 | (__v8hi)_mm_unpackhi_epi16(__A, __B), |
||
| 1684 | (__v8hi) _mm_setzero_si128()); |
||
| 1685 | } |
||
| 1686 | |||
| 1687 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1688 | _mm256_mask_unpackhi_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1689 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1690 | (__v16hi)_mm256_unpackhi_epi16(__A, __B), |
||
| 1691 | (__v16hi)__W); |
||
| 1692 | } |
||
| 1693 | |||
| 1694 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1695 | _mm256_maskz_unpackhi_epi16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1696 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1697 | (__v16hi)_mm256_unpackhi_epi16(__A, __B), |
||
| 1698 | (__v16hi)_mm256_setzero_si256()); |
||
| 1699 | } |
||
| 1700 | |||
| 1701 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1702 | _mm_mask_unpacklo_epi8(__m128i __W, __mmask16 __U, __m128i __A, __m128i __B) { |
||
| 1703 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1704 | (__v16qi)_mm_unpacklo_epi8(__A, __B), |
||
| 1705 | (__v16qi)__W); |
||
| 1706 | } |
||
| 1707 | |||
| 1708 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1709 | _mm_maskz_unpacklo_epi8(__mmask16 __U, __m128i __A, __m128i __B) { |
||
| 1710 | return (__m128i)__builtin_ia32_selectb_128((__mmask16)__U, |
||
| 1711 | (__v16qi)_mm_unpacklo_epi8(__A, __B), |
||
| 1712 | (__v16qi)_mm_setzero_si128()); |
||
| 1713 | } |
||
| 1714 | |||
| 1715 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1716 | _mm256_mask_unpacklo_epi8(__m256i __W, __mmask32 __U, __m256i __A, __m256i __B) { |
||
| 1717 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1718 | (__v32qi)_mm256_unpacklo_epi8(__A, __B), |
||
| 1719 | (__v32qi)__W); |
||
| 1720 | } |
||
| 1721 | |||
| 1722 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1723 | _mm256_maskz_unpacklo_epi8(__mmask32 __U, __m256i __A, __m256i __B) { |
||
| 1724 | return (__m256i)__builtin_ia32_selectb_256((__mmask32)__U, |
||
| 1725 | (__v32qi)_mm256_unpacklo_epi8(__A, __B), |
||
| 1726 | (__v32qi)_mm256_setzero_si256()); |
||
| 1727 | } |
||
| 1728 | |||
| 1729 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1730 | _mm_mask_unpacklo_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1731 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1732 | (__v8hi)_mm_unpacklo_epi16(__A, __B), |
||
| 1733 | (__v8hi)__W); |
||
| 1734 | } |
||
| 1735 | |||
| 1736 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1737 | _mm_maskz_unpacklo_epi16(__mmask8 __U, __m128i __A, __m128i __B) { |
||
| 1738 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1739 | (__v8hi)_mm_unpacklo_epi16(__A, __B), |
||
| 1740 | (__v8hi) _mm_setzero_si128()); |
||
| 1741 | } |
||
| 1742 | |||
| 1743 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1744 | _mm256_mask_unpacklo_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1745 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1746 | (__v16hi)_mm256_unpacklo_epi16(__A, __B), |
||
| 1747 | (__v16hi)__W); |
||
| 1748 | } |
||
| 1749 | |||
| 1750 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1751 | _mm256_maskz_unpacklo_epi16(__mmask16 __U, __m256i __A, __m256i __B) { |
||
| 1752 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1753 | (__v16hi)_mm256_unpacklo_epi16(__A, __B), |
||
| 1754 | (__v16hi)_mm256_setzero_si256()); |
||
| 1755 | } |
||
| 1756 | |||
| 1757 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1758 | _mm_mask_cvtepi8_epi16(__m128i __W, __mmask8 __U, __m128i __A) |
||
| 1759 | { |
||
| 1760 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1761 | (__v8hi)_mm_cvtepi8_epi16(__A), |
||
| 1762 | (__v8hi)__W); |
||
| 1763 | } |
||
| 1764 | |||
| 1765 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1766 | _mm_maskz_cvtepi8_epi16(__mmask8 __U, __m128i __A) |
||
| 1767 | { |
||
| 1768 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1769 | (__v8hi)_mm_cvtepi8_epi16(__A), |
||
| 1770 | (__v8hi)_mm_setzero_si128()); |
||
| 1771 | } |
||
| 1772 | |||
| 1773 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1774 | _mm256_mask_cvtepi8_epi16(__m256i __W, __mmask16 __U, __m128i __A) |
||
| 1775 | { |
||
| 1776 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1777 | (__v16hi)_mm256_cvtepi8_epi16(__A), |
||
| 1778 | (__v16hi)__W); |
||
| 1779 | } |
||
| 1780 | |||
| 1781 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1782 | _mm256_maskz_cvtepi8_epi16(__mmask16 __U, __m128i __A) |
||
| 1783 | { |
||
| 1784 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1785 | (__v16hi)_mm256_cvtepi8_epi16(__A), |
||
| 1786 | (__v16hi)_mm256_setzero_si256()); |
||
| 1787 | } |
||
| 1788 | |||
| 1789 | |||
| 1790 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1791 | _mm_mask_cvtepu8_epi16(__m128i __W, __mmask8 __U, __m128i __A) |
||
| 1792 | { |
||
| 1793 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1794 | (__v8hi)_mm_cvtepu8_epi16(__A), |
||
| 1795 | (__v8hi)__W); |
||
| 1796 | } |
||
| 1797 | |||
| 1798 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1799 | _mm_maskz_cvtepu8_epi16(__mmask8 __U, __m128i __A) |
||
| 1800 | { |
||
| 1801 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1802 | (__v8hi)_mm_cvtepu8_epi16(__A), |
||
| 1803 | (__v8hi)_mm_setzero_si128()); |
||
| 1804 | } |
||
| 1805 | |||
| 1806 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1807 | _mm256_mask_cvtepu8_epi16(__m256i __W, __mmask16 __U, __m128i __A) |
||
| 1808 | { |
||
| 1809 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1810 | (__v16hi)_mm256_cvtepu8_epi16(__A), |
||
| 1811 | (__v16hi)__W); |
||
| 1812 | } |
||
| 1813 | |||
| 1814 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1815 | _mm256_maskz_cvtepu8_epi16 (__mmask16 __U, __m128i __A) |
||
| 1816 | { |
||
| 1817 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1818 | (__v16hi)_mm256_cvtepu8_epi16(__A), |
||
| 1819 | (__v16hi)_mm256_setzero_si256()); |
||
| 1820 | } |
||
| 1821 | |||
| 1822 | |||
| 1823 | #define _mm_mask_shufflehi_epi16(W, U, A, imm) \ |
||
| 1824 | ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \ |
||
| 1825 | (__v8hi)_mm_shufflehi_epi16((A), (imm)), \ |
||
| 1826 | (__v8hi)(__m128i)(W))) |
||
| 1827 | |||
| 1828 | #define _mm_maskz_shufflehi_epi16(U, A, imm) \ |
||
| 1829 | ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \ |
||
| 1830 | (__v8hi)_mm_shufflehi_epi16((A), (imm)), \ |
||
| 1831 | (__v8hi)_mm_setzero_si128())) |
||
| 1832 | |||
| 1833 | #define _mm256_mask_shufflehi_epi16(W, U, A, imm) \ |
||
| 1834 | ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \ |
||
| 1835 | (__v16hi)_mm256_shufflehi_epi16((A), (imm)), \ |
||
| 1836 | (__v16hi)(__m256i)(W))) |
||
| 1837 | |||
| 1838 | #define _mm256_maskz_shufflehi_epi16(U, A, imm) \ |
||
| 1839 | ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \ |
||
| 1840 | (__v16hi)_mm256_shufflehi_epi16((A), (imm)), \ |
||
| 1841 | (__v16hi)_mm256_setzero_si256())) |
||
| 1842 | |||
| 1843 | #define _mm_mask_shufflelo_epi16(W, U, A, imm) \ |
||
| 1844 | ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \ |
||
| 1845 | (__v8hi)_mm_shufflelo_epi16((A), (imm)), \ |
||
| 1846 | (__v8hi)(__m128i)(W))) |
||
| 1847 | |||
| 1848 | #define _mm_maskz_shufflelo_epi16(U, A, imm) \ |
||
| 1849 | ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \ |
||
| 1850 | (__v8hi)_mm_shufflelo_epi16((A), (imm)), \ |
||
| 1851 | (__v8hi)_mm_setzero_si128())) |
||
| 1852 | |||
| 1853 | #define _mm256_mask_shufflelo_epi16(W, U, A, imm) \ |
||
| 1854 | ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \ |
||
| 1855 | (__v16hi)_mm256_shufflelo_epi16((A), \ |
||
| 1856 | (imm)), \ |
||
| 1857 | (__v16hi)(__m256i)(W))) |
||
| 1858 | |||
| 1859 | #define _mm256_maskz_shufflelo_epi16(U, A, imm) \ |
||
| 1860 | ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \ |
||
| 1861 | (__v16hi)_mm256_shufflelo_epi16((A), \ |
||
| 1862 | (imm)), \ |
||
| 1863 | (__v16hi)_mm256_setzero_si256())) |
||
| 1864 | |||
| 1865 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1866 | _mm256_sllv_epi16(__m256i __A, __m256i __B) |
||
| 1867 | { |
||
| 1868 | return (__m256i)__builtin_ia32_psllv16hi((__v16hi)__A, (__v16hi)__B); |
||
| 1869 | } |
||
| 1870 | |||
| 1871 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1872 | _mm256_mask_sllv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 1873 | { |
||
| 1874 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1875 | (__v16hi)_mm256_sllv_epi16(__A, __B), |
||
| 1876 | (__v16hi)__W); |
||
| 1877 | } |
||
| 1878 | |||
| 1879 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1880 | _mm256_maskz_sllv_epi16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 1881 | { |
||
| 1882 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1883 | (__v16hi)_mm256_sllv_epi16(__A, __B), |
||
| 1884 | (__v16hi)_mm256_setzero_si256()); |
||
| 1885 | } |
||
| 1886 | |||
| 1887 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1888 | _mm_sllv_epi16(__m128i __A, __m128i __B) |
||
| 1889 | { |
||
| 1890 | return (__m128i)__builtin_ia32_psllv8hi((__v8hi)__A, (__v8hi)__B); |
||
| 1891 | } |
||
| 1892 | |||
| 1893 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1894 | _mm_mask_sllv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 1895 | { |
||
| 1896 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1897 | (__v8hi)_mm_sllv_epi16(__A, __B), |
||
| 1898 | (__v8hi)__W); |
||
| 1899 | } |
||
| 1900 | |||
| 1901 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1902 | _mm_maskz_sllv_epi16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 1903 | { |
||
| 1904 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1905 | (__v8hi)_mm_sllv_epi16(__A, __B), |
||
| 1906 | (__v8hi)_mm_setzero_si128()); |
||
| 1907 | } |
||
| 1908 | |||
| 1909 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1910 | _mm_mask_sll_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 1911 | { |
||
| 1912 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1913 | (__v8hi)_mm_sll_epi16(__A, __B), |
||
| 1914 | (__v8hi)__W); |
||
| 1915 | } |
||
| 1916 | |||
| 1917 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1918 | _mm_maskz_sll_epi16 (__mmask8 __U, __m128i __A, __m128i __B) |
||
| 1919 | { |
||
| 1920 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1921 | (__v8hi)_mm_sll_epi16(__A, __B), |
||
| 1922 | (__v8hi)_mm_setzero_si128()); |
||
| 1923 | } |
||
| 1924 | |||
| 1925 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1926 | _mm256_mask_sll_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B) |
||
| 1927 | { |
||
| 1928 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1929 | (__v16hi)_mm256_sll_epi16(__A, __B), |
||
| 1930 | (__v16hi)__W); |
||
| 1931 | } |
||
| 1932 | |||
| 1933 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1934 | _mm256_maskz_sll_epi16(__mmask16 __U, __m256i __A, __m128i __B) |
||
| 1935 | { |
||
| 1936 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1937 | (__v16hi)_mm256_sll_epi16(__A, __B), |
||
| 1938 | (__v16hi)_mm256_setzero_si256()); |
||
| 1939 | } |
||
| 1940 | |||
| 1941 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1942 | _mm_mask_slli_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B) |
||
| 1943 | { |
||
| 1944 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1945 | (__v8hi)_mm_slli_epi16(__A, (int)__B), |
||
| 1946 | (__v8hi)__W); |
||
| 1947 | } |
||
| 1948 | |||
| 1949 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1950 | _mm_maskz_slli_epi16 (__mmask8 __U, __m128i __A, unsigned int __B) |
||
| 1951 | { |
||
| 1952 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 1953 | (__v8hi)_mm_slli_epi16(__A, (int)__B), |
||
| 1954 | (__v8hi)_mm_setzero_si128()); |
||
| 1955 | } |
||
| 1956 | |||
| 1957 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1958 | _mm256_mask_slli_epi16(__m256i __W, __mmask16 __U, __m256i __A, |
||
| 1959 | unsigned int __B) |
||
| 1960 | { |
||
| 1961 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1962 | (__v16hi)_mm256_slli_epi16(__A, (int)__B), |
||
| 1963 | (__v16hi)__W); |
||
| 1964 | } |
||
| 1965 | |||
| 1966 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1967 | _mm256_maskz_slli_epi16(__mmask16 __U, __m256i __A, unsigned int __B) |
||
| 1968 | { |
||
| 1969 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1970 | (__v16hi)_mm256_slli_epi16(__A, (int)__B), |
||
| 1971 | (__v16hi)_mm256_setzero_si256()); |
||
| 1972 | } |
||
| 1973 | |||
| 1974 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1975 | _mm256_srlv_epi16(__m256i __A, __m256i __B) |
||
| 1976 | { |
||
| 1977 | return (__m256i)__builtin_ia32_psrlv16hi((__v16hi)__A, (__v16hi)__B); |
||
| 1978 | } |
||
| 1979 | |||
| 1980 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1981 | _mm256_mask_srlv_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 1982 | { |
||
| 1983 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1984 | (__v16hi)_mm256_srlv_epi16(__A, __B), |
||
| 1985 | (__v16hi)__W); |
||
| 1986 | } |
||
| 1987 | |||
| 1988 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 1989 | _mm256_maskz_srlv_epi16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 1990 | { |
||
| 1991 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 1992 | (__v16hi)_mm256_srlv_epi16(__A, __B), |
||
| 1993 | (__v16hi)_mm256_setzero_si256()); |
||
| 1994 | } |
||
| 1995 | |||
| 1996 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 1997 | _mm_srlv_epi16(__m128i __A, __m128i __B) |
||
| 1998 | { |
||
| 1999 | return (__m128i)__builtin_ia32_psrlv8hi((__v8hi)__A, (__v8hi)__B); |
||
| 2000 | } |
||
| 2001 | |||
| 2002 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2003 | _mm_mask_srlv_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 2004 | { |
||
| 2005 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2006 | (__v8hi)_mm_srlv_epi16(__A, __B), |
||
| 2007 | (__v8hi)__W); |
||
| 2008 | } |
||
| 2009 | |||
| 2010 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2011 | _mm_maskz_srlv_epi16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 2012 | { |
||
| 2013 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2014 | (__v8hi)_mm_srlv_epi16(__A, __B), |
||
| 2015 | (__v8hi)_mm_setzero_si128()); |
||
| 2016 | } |
||
| 2017 | |||
| 2018 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2019 | _mm256_srav_epi16(__m256i __A, __m256i __B) |
||
| 2020 | { |
||
| 2021 | return (__m256i)__builtin_ia32_psrav16hi((__v16hi)__A, (__v16hi)__B); |
||
| 2022 | } |
||
| 2023 | |||
| 2024 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2025 | _mm256_mask_srav_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m256i __B) |
||
| 2026 | { |
||
| 2027 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2028 | (__v16hi)_mm256_srav_epi16(__A, __B), |
||
| 2029 | (__v16hi)__W); |
||
| 2030 | } |
||
| 2031 | |||
| 2032 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2033 | _mm256_maskz_srav_epi16(__mmask16 __U, __m256i __A, __m256i __B) |
||
| 2034 | { |
||
| 2035 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2036 | (__v16hi)_mm256_srav_epi16(__A, __B), |
||
| 2037 | (__v16hi)_mm256_setzero_si256()); |
||
| 2038 | } |
||
| 2039 | |||
| 2040 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2041 | _mm_srav_epi16(__m128i __A, __m128i __B) |
||
| 2042 | { |
||
| 2043 | return (__m128i)__builtin_ia32_psrav8hi((__v8hi)__A, (__v8hi)__B); |
||
| 2044 | } |
||
| 2045 | |||
| 2046 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2047 | _mm_mask_srav_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 2048 | { |
||
| 2049 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2050 | (__v8hi)_mm_srav_epi16(__A, __B), |
||
| 2051 | (__v8hi)__W); |
||
| 2052 | } |
||
| 2053 | |||
| 2054 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2055 | _mm_maskz_srav_epi16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 2056 | { |
||
| 2057 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2058 | (__v8hi)_mm_srav_epi16(__A, __B), |
||
| 2059 | (__v8hi)_mm_setzero_si128()); |
||
| 2060 | } |
||
| 2061 | |||
| 2062 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2063 | _mm_mask_sra_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 2064 | { |
||
| 2065 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2066 | (__v8hi)_mm_sra_epi16(__A, __B), |
||
| 2067 | (__v8hi)__W); |
||
| 2068 | } |
||
| 2069 | |||
| 2070 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2071 | _mm_maskz_sra_epi16(__mmask8 __U, __m128i __A, __m128i __B) |
||
| 2072 | { |
||
| 2073 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2074 | (__v8hi)_mm_sra_epi16(__A, __B), |
||
| 2075 | (__v8hi)_mm_setzero_si128()); |
||
| 2076 | } |
||
| 2077 | |||
| 2078 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2079 | _mm256_mask_sra_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B) |
||
| 2080 | { |
||
| 2081 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2082 | (__v16hi)_mm256_sra_epi16(__A, __B), |
||
| 2083 | (__v16hi)__W); |
||
| 2084 | } |
||
| 2085 | |||
| 2086 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2087 | _mm256_maskz_sra_epi16(__mmask16 __U, __m256i __A, __m128i __B) |
||
| 2088 | { |
||
| 2089 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2090 | (__v16hi)_mm256_sra_epi16(__A, __B), |
||
| 2091 | (__v16hi)_mm256_setzero_si256()); |
||
| 2092 | } |
||
| 2093 | |||
| 2094 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2095 | _mm_mask_srai_epi16(__m128i __W, __mmask8 __U, __m128i __A, unsigned int __B) |
||
| 2096 | { |
||
| 2097 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2098 | (__v8hi)_mm_srai_epi16(__A, (int)__B), |
||
| 2099 | (__v8hi)__W); |
||
| 2100 | } |
||
| 2101 | |||
| 2102 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2103 | _mm_maskz_srai_epi16(__mmask8 __U, __m128i __A, unsigned int __B) |
||
| 2104 | { |
||
| 2105 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2106 | (__v8hi)_mm_srai_epi16(__A, (int)__B), |
||
| 2107 | (__v8hi)_mm_setzero_si128()); |
||
| 2108 | } |
||
| 2109 | |||
| 2110 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2111 | _mm256_mask_srai_epi16(__m256i __W, __mmask16 __U, __m256i __A, |
||
| 2112 | unsigned int __B) |
||
| 2113 | { |
||
| 2114 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2115 | (__v16hi)_mm256_srai_epi16(__A, (int)__B), |
||
| 2116 | (__v16hi)__W); |
||
| 2117 | } |
||
| 2118 | |||
| 2119 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2120 | _mm256_maskz_srai_epi16(__mmask16 __U, __m256i __A, unsigned int __B) |
||
| 2121 | { |
||
| 2122 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2123 | (__v16hi)_mm256_srai_epi16(__A, (int)__B), |
||
| 2124 | (__v16hi)_mm256_setzero_si256()); |
||
| 2125 | } |
||
| 2126 | |||
| 2127 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2128 | _mm_mask_srl_epi16(__m128i __W, __mmask8 __U, __m128i __A, __m128i __B) |
||
| 2129 | { |
||
| 2130 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2131 | (__v8hi)_mm_srl_epi16(__A, __B), |
||
| 2132 | (__v8hi)__W); |
||
| 2133 | } |
||
| 2134 | |||
| 2135 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2136 | _mm_maskz_srl_epi16 (__mmask8 __U, __m128i __A, __m128i __B) |
||
| 2137 | { |
||
| 2138 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2139 | (__v8hi)_mm_srl_epi16(__A, __B), |
||
| 2140 | (__v8hi)_mm_setzero_si128()); |
||
| 2141 | } |
||
| 2142 | |||
| 2143 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2144 | _mm256_mask_srl_epi16(__m256i __W, __mmask16 __U, __m256i __A, __m128i __B) |
||
| 2145 | { |
||
| 2146 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2147 | (__v16hi)_mm256_srl_epi16(__A, __B), |
||
| 2148 | (__v16hi)__W); |
||
| 2149 | } |
||
| 2150 | |||
| 2151 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2152 | _mm256_maskz_srl_epi16(__mmask16 __U, __m256i __A, __m128i __B) |
||
| 2153 | { |
||
| 2154 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2155 | (__v16hi)_mm256_srl_epi16(__A, __B), |
||
| 2156 | (__v16hi)_mm256_setzero_si256()); |
||
| 2157 | } |
||
| 2158 | |||
| 2159 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2160 | _mm_mask_srli_epi16(__m128i __W, __mmask8 __U, __m128i __A, int __B) |
||
| 2161 | { |
||
| 2162 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2163 | (__v8hi)_mm_srli_epi16(__A, __B), |
||
| 2164 | (__v8hi)__W); |
||
| 2165 | } |
||
| 2166 | |||
| 2167 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2168 | _mm_maskz_srli_epi16 (__mmask8 __U, __m128i __A, int __B) |
||
| 2169 | { |
||
| 2170 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__U, |
||
| 2171 | (__v8hi)_mm_srli_epi16(__A, __B), |
||
| 2172 | (__v8hi)_mm_setzero_si128()); |
||
| 2173 | } |
||
| 2174 | |||
| 2175 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2176 | _mm256_mask_srli_epi16(__m256i __W, __mmask16 __U, __m256i __A, int __B) |
||
| 2177 | { |
||
| 2178 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2179 | (__v16hi)_mm256_srli_epi16(__A, __B), |
||
| 2180 | (__v16hi)__W); |
||
| 2181 | } |
||
| 2182 | |||
| 2183 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2184 | _mm256_maskz_srli_epi16(__mmask16 __U, __m256i __A, int __B) |
||
| 2185 | { |
||
| 2186 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__U, |
||
| 2187 | (__v16hi)_mm256_srli_epi16(__A, __B), |
||
| 2188 | (__v16hi)_mm256_setzero_si256()); |
||
| 2189 | } |
||
| 2190 | |||
| 2191 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2192 | _mm_mask_mov_epi16 (__m128i __W, __mmask8 __U, __m128i __A) |
||
| 2193 | { |
||
| 2194 | return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U, |
||
| 2195 | (__v8hi) __A, |
||
| 2196 | (__v8hi) __W); |
||
| 2197 | } |
||
| 2198 | |||
| 2199 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2200 | _mm_maskz_mov_epi16 (__mmask8 __U, __m128i __A) |
||
| 2201 | { |
||
| 2202 | return (__m128i) __builtin_ia32_selectw_128 ((__mmask8) __U, |
||
| 2203 | (__v8hi) __A, |
||
| 2204 | (__v8hi) _mm_setzero_si128 ()); |
||
| 2205 | } |
||
| 2206 | |||
| 2207 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2208 | _mm256_mask_mov_epi16 (__m256i __W, __mmask16 __U, __m256i __A) |
||
| 2209 | { |
||
| 2210 | return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U, |
||
| 2211 | (__v16hi) __A, |
||
| 2212 | (__v16hi) __W); |
||
| 2213 | } |
||
| 2214 | |||
| 2215 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2216 | _mm256_maskz_mov_epi16 (__mmask16 __U, __m256i __A) |
||
| 2217 | { |
||
| 2218 | return (__m256i) __builtin_ia32_selectw_256 ((__mmask16) __U, |
||
| 2219 | (__v16hi) __A, |
||
| 2220 | (__v16hi) _mm256_setzero_si256 ()); |
||
| 2221 | } |
||
| 2222 | |||
| 2223 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2224 | _mm_mask_mov_epi8 (__m128i __W, __mmask16 __U, __m128i __A) |
||
| 2225 | { |
||
| 2226 | return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U, |
||
| 2227 | (__v16qi) __A, |
||
| 2228 | (__v16qi) __W); |
||
| 2229 | } |
||
| 2230 | |||
| 2231 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2232 | _mm_maskz_mov_epi8 (__mmask16 __U, __m128i __A) |
||
| 2233 | { |
||
| 2234 | return (__m128i) __builtin_ia32_selectb_128 ((__mmask16) __U, |
||
| 2235 | (__v16qi) __A, |
||
| 2236 | (__v16qi) _mm_setzero_si128 ()); |
||
| 2237 | } |
||
| 2238 | |||
| 2239 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2240 | _mm256_mask_mov_epi8 (__m256i __W, __mmask32 __U, __m256i __A) |
||
| 2241 | { |
||
| 2242 | return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U, |
||
| 2243 | (__v32qi) __A, |
||
| 2244 | (__v32qi) __W); |
||
| 2245 | } |
||
| 2246 | |||
| 2247 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2248 | _mm256_maskz_mov_epi8 (__mmask32 __U, __m256i __A) |
||
| 2249 | { |
||
| 2250 | return (__m256i) __builtin_ia32_selectb_256 ((__mmask32) __U, |
||
| 2251 | (__v32qi) __A, |
||
| 2252 | (__v32qi) _mm256_setzero_si256 ()); |
||
| 2253 | } |
||
| 2254 | |||
| 2255 | |||
| 2256 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2257 | _mm_mask_set1_epi8 (__m128i __O, __mmask16 __M, char __A) |
||
| 2258 | { |
||
| 2259 | return (__m128i) __builtin_ia32_selectb_128(__M, |
||
| 2260 | (__v16qi) _mm_set1_epi8(__A), |
||
| 2261 | (__v16qi) __O); |
||
| 2262 | } |
||
| 2263 | |||
| 2264 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2265 | _mm_maskz_set1_epi8 (__mmask16 __M, char __A) |
||
| 2266 | { |
||
| 2267 | return (__m128i) __builtin_ia32_selectb_128(__M, |
||
| 2268 | (__v16qi) _mm_set1_epi8(__A), |
||
| 2269 | (__v16qi) _mm_setzero_si128()); |
||
| 2270 | } |
||
| 2271 | |||
| 2272 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2273 | _mm256_mask_set1_epi8 (__m256i __O, __mmask32 __M, char __A) |
||
| 2274 | { |
||
| 2275 | return (__m256i) __builtin_ia32_selectb_256(__M, |
||
| 2276 | (__v32qi) _mm256_set1_epi8(__A), |
||
| 2277 | (__v32qi) __O); |
||
| 2278 | } |
||
| 2279 | |||
| 2280 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2281 | _mm256_maskz_set1_epi8 (__mmask32 __M, char __A) |
||
| 2282 | { |
||
| 2283 | return (__m256i) __builtin_ia32_selectb_256(__M, |
||
| 2284 | (__v32qi) _mm256_set1_epi8(__A), |
||
| 2285 | (__v32qi) _mm256_setzero_si256()); |
||
| 2286 | } |
||
| 2287 | |||
| 2288 | static __inline __m128i __DEFAULT_FN_ATTRS128 |
||
| 2289 | _mm_loadu_epi16 (void const *__P) |
||
| 2290 | { |
||
| 2291 | struct __loadu_epi16 { |
||
| 2292 | __m128i_u __v; |
||
| 2293 | } __attribute__((__packed__, __may_alias__)); |
||
| 2294 | return ((const struct __loadu_epi16*)__P)->__v; |
||
| 2295 | } |
||
| 2296 | |||
| 2297 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2298 | _mm_mask_loadu_epi16 (__m128i __W, __mmask8 __U, void const *__P) |
||
| 2299 | { |
||
| 2300 | return (__m128i) __builtin_ia32_loaddquhi128_mask ((const __v8hi *) __P, |
||
| 2301 | (__v8hi) __W, |
||
| 2302 | (__mmask8) __U); |
||
| 2303 | } |
||
| 2304 | |||
| 2305 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2306 | _mm_maskz_loadu_epi16 (__mmask8 __U, void const *__P) |
||
| 2307 | { |
||
| 2308 | return (__m128i) __builtin_ia32_loaddquhi128_mask ((const __v8hi *) __P, |
||
| 2309 | (__v8hi) |
||
| 2310 | _mm_setzero_si128 (), |
||
| 2311 | (__mmask8) __U); |
||
| 2312 | } |
||
| 2313 | |||
| 2314 | static __inline __m256i __DEFAULT_FN_ATTRS256 |
||
| 2315 | _mm256_loadu_epi16 (void const *__P) |
||
| 2316 | { |
||
| 2317 | struct __loadu_epi16 { |
||
| 2318 | __m256i_u __v; |
||
| 2319 | } __attribute__((__packed__, __may_alias__)); |
||
| 2320 | return ((const struct __loadu_epi16*)__P)->__v; |
||
| 2321 | } |
||
| 2322 | |||
| 2323 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2324 | _mm256_mask_loadu_epi16 (__m256i __W, __mmask16 __U, void const *__P) |
||
| 2325 | { |
||
| 2326 | return (__m256i) __builtin_ia32_loaddquhi256_mask ((const __v16hi *) __P, |
||
| 2327 | (__v16hi) __W, |
||
| 2328 | (__mmask16) __U); |
||
| 2329 | } |
||
| 2330 | |||
| 2331 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2332 | _mm256_maskz_loadu_epi16 (__mmask16 __U, void const *__P) |
||
| 2333 | { |
||
| 2334 | return (__m256i) __builtin_ia32_loaddquhi256_mask ((const __v16hi *) __P, |
||
| 2335 | (__v16hi) |
||
| 2336 | _mm256_setzero_si256 (), |
||
| 2337 | (__mmask16) __U); |
||
| 2338 | } |
||
| 2339 | |||
| 2340 | static __inline __m128i __DEFAULT_FN_ATTRS128 |
||
| 2341 | _mm_loadu_epi8 (void const *__P) |
||
| 2342 | { |
||
| 2343 | struct __loadu_epi8 { |
||
| 2344 | __m128i_u __v; |
||
| 2345 | } __attribute__((__packed__, __may_alias__)); |
||
| 2346 | return ((const struct __loadu_epi8*)__P)->__v; |
||
| 2347 | } |
||
| 2348 | |||
| 2349 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2350 | _mm_mask_loadu_epi8 (__m128i __W, __mmask16 __U, void const *__P) |
||
| 2351 | { |
||
| 2352 | return (__m128i) __builtin_ia32_loaddquqi128_mask ((const __v16qi *) __P, |
||
| 2353 | (__v16qi) __W, |
||
| 2354 | (__mmask16) __U); |
||
| 2355 | } |
||
| 2356 | |||
| 2357 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2358 | _mm_maskz_loadu_epi8 (__mmask16 __U, void const *__P) |
||
| 2359 | { |
||
| 2360 | return (__m128i) __builtin_ia32_loaddquqi128_mask ((const __v16qi *) __P, |
||
| 2361 | (__v16qi) |
||
| 2362 | _mm_setzero_si128 (), |
||
| 2363 | (__mmask16) __U); |
||
| 2364 | } |
||
| 2365 | |||
| 2366 | static __inline __m256i __DEFAULT_FN_ATTRS256 |
||
| 2367 | _mm256_loadu_epi8 (void const *__P) |
||
| 2368 | { |
||
| 2369 | struct __loadu_epi8 { |
||
| 2370 | __m256i_u __v; |
||
| 2371 | } __attribute__((__packed__, __may_alias__)); |
||
| 2372 | return ((const struct __loadu_epi8*)__P)->__v; |
||
| 2373 | } |
||
| 2374 | |||
| 2375 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2376 | _mm256_mask_loadu_epi8 (__m256i __W, __mmask32 __U, void const *__P) |
||
| 2377 | { |
||
| 2378 | return (__m256i) __builtin_ia32_loaddquqi256_mask ((const __v32qi *) __P, |
||
| 2379 | (__v32qi) __W, |
||
| 2380 | (__mmask32) __U); |
||
| 2381 | } |
||
| 2382 | |||
| 2383 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2384 | _mm256_maskz_loadu_epi8 (__mmask32 __U, void const *__P) |
||
| 2385 | { |
||
| 2386 | return (__m256i) __builtin_ia32_loaddquqi256_mask ((const __v32qi *) __P, |
||
| 2387 | (__v32qi) |
||
| 2388 | _mm256_setzero_si256 (), |
||
| 2389 | (__mmask32) __U); |
||
| 2390 | } |
||
| 2391 | |||
| 2392 | static __inline void __DEFAULT_FN_ATTRS128 |
||
| 2393 | _mm_storeu_epi16 (void *__P, __m128i __A) |
||
| 2394 | { |
||
| 2395 | struct __storeu_epi16 { |
||
| 2396 | __m128i_u __v; |
||
| 2397 | } __attribute__((__packed__, __may_alias__)); |
||
| 2398 | ((struct __storeu_epi16*)__P)->__v = __A; |
||
| 2399 | } |
||
| 2400 | |||
| 2401 | static __inline__ void __DEFAULT_FN_ATTRS128 |
||
| 2402 | _mm_mask_storeu_epi16 (void *__P, __mmask8 __U, __m128i __A) |
||
| 2403 | { |
||
| 2404 | __builtin_ia32_storedquhi128_mask ((__v8hi *) __P, |
||
| 2405 | (__v8hi) __A, |
||
| 2406 | (__mmask8) __U); |
||
| 2407 | } |
||
| 2408 | |||
| 2409 | static __inline void __DEFAULT_FN_ATTRS256 |
||
| 2410 | _mm256_storeu_epi16 (void *__P, __m256i __A) |
||
| 2411 | { |
||
| 2412 | struct __storeu_epi16 { |
||
| 2413 | __m256i_u __v; |
||
| 2414 | } __attribute__((__packed__, __may_alias__)); |
||
| 2415 | ((struct __storeu_epi16*)__P)->__v = __A; |
||
| 2416 | } |
||
| 2417 | |||
| 2418 | static __inline__ void __DEFAULT_FN_ATTRS256 |
||
| 2419 | _mm256_mask_storeu_epi16 (void *__P, __mmask16 __U, __m256i __A) |
||
| 2420 | { |
||
| 2421 | __builtin_ia32_storedquhi256_mask ((__v16hi *) __P, |
||
| 2422 | (__v16hi) __A, |
||
| 2423 | (__mmask16) __U); |
||
| 2424 | } |
||
| 2425 | |||
| 2426 | static __inline void __DEFAULT_FN_ATTRS128 |
||
| 2427 | _mm_storeu_epi8 (void *__P, __m128i __A) |
||
| 2428 | { |
||
| 2429 | struct __storeu_epi8 { |
||
| 2430 | __m128i_u __v; |
||
| 2431 | } __attribute__((__packed__, __may_alias__)); |
||
| 2432 | ((struct __storeu_epi8*)__P)->__v = __A; |
||
| 2433 | } |
||
| 2434 | |||
| 2435 | static __inline__ void __DEFAULT_FN_ATTRS128 |
||
| 2436 | _mm_mask_storeu_epi8 (void *__P, __mmask16 __U, __m128i __A) |
||
| 2437 | { |
||
| 2438 | __builtin_ia32_storedquqi128_mask ((__v16qi *) __P, |
||
| 2439 | (__v16qi) __A, |
||
| 2440 | (__mmask16) __U); |
||
| 2441 | } |
||
| 2442 | |||
| 2443 | static __inline void __DEFAULT_FN_ATTRS256 |
||
| 2444 | _mm256_storeu_epi8 (void *__P, __m256i __A) |
||
| 2445 | { |
||
| 2446 | struct __storeu_epi8 { |
||
| 2447 | __m256i_u __v; |
||
| 2448 | } __attribute__((__packed__, __may_alias__)); |
||
| 2449 | ((struct __storeu_epi8*)__P)->__v = __A; |
||
| 2450 | } |
||
| 2451 | |||
| 2452 | static __inline__ void __DEFAULT_FN_ATTRS256 |
||
| 2453 | _mm256_mask_storeu_epi8 (void *__P, __mmask32 __U, __m256i __A) |
||
| 2454 | { |
||
| 2455 | __builtin_ia32_storedquqi256_mask ((__v32qi *) __P, |
||
| 2456 | (__v32qi) __A, |
||
| 2457 | (__mmask32) __U); |
||
| 2458 | } |
||
| 2459 | |||
| 2460 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 |
||
| 2461 | _mm_test_epi8_mask (__m128i __A, __m128i __B) |
||
| 2462 | { |
||
| 2463 | return _mm_cmpneq_epi8_mask (_mm_and_si128(__A, __B), _mm_setzero_si128()); |
||
| 2464 | } |
||
| 2465 | |||
| 2466 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 |
||
| 2467 | _mm_mask_test_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B) |
||
| 2468 | { |
||
| 2469 | return _mm_mask_cmpneq_epi8_mask (__U, _mm_and_si128 (__A, __B), |
||
| 2470 | _mm_setzero_si128()); |
||
| 2471 | } |
||
| 2472 | |||
| 2473 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 |
||
| 2474 | _mm256_test_epi8_mask (__m256i __A, __m256i __B) |
||
| 2475 | { |
||
| 2476 | return _mm256_cmpneq_epi8_mask (_mm256_and_si256(__A, __B), |
||
| 2477 | _mm256_setzero_si256()); |
||
| 2478 | } |
||
| 2479 | |||
| 2480 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 |
||
| 2481 | _mm256_mask_test_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B) |
||
| 2482 | { |
||
| 2483 | return _mm256_mask_cmpneq_epi8_mask (__U, _mm256_and_si256(__A, __B), |
||
| 2484 | _mm256_setzero_si256()); |
||
| 2485 | } |
||
| 2486 | |||
| 2487 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 |
||
| 2488 | _mm_test_epi16_mask (__m128i __A, __m128i __B) |
||
| 2489 | { |
||
| 2490 | return _mm_cmpneq_epi16_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
||
| 2491 | } |
||
| 2492 | |||
| 2493 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 |
||
| 2494 | _mm_mask_test_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B) |
||
| 2495 | { |
||
| 2496 | return _mm_mask_cmpneq_epi16_mask (__U, _mm_and_si128 (__A, __B), |
||
| 2497 | _mm_setzero_si128()); |
||
| 2498 | } |
||
| 2499 | |||
| 2500 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 |
||
| 2501 | _mm256_test_epi16_mask (__m256i __A, __m256i __B) |
||
| 2502 | { |
||
| 2503 | return _mm256_cmpneq_epi16_mask (_mm256_and_si256 (__A, __B), |
||
| 2504 | _mm256_setzero_si256 ()); |
||
| 2505 | } |
||
| 2506 | |||
| 2507 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 |
||
| 2508 | _mm256_mask_test_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B) |
||
| 2509 | { |
||
| 2510 | return _mm256_mask_cmpneq_epi16_mask (__U, _mm256_and_si256(__A, __B), |
||
| 2511 | _mm256_setzero_si256()); |
||
| 2512 | } |
||
| 2513 | |||
| 2514 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 |
||
| 2515 | _mm_testn_epi8_mask (__m128i __A, __m128i __B) |
||
| 2516 | { |
||
| 2517 | return _mm_cmpeq_epi8_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
||
| 2518 | } |
||
| 2519 | |||
| 2520 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 |
||
| 2521 | _mm_mask_testn_epi8_mask (__mmask16 __U, __m128i __A, __m128i __B) |
||
| 2522 | { |
||
| 2523 | return _mm_mask_cmpeq_epi8_mask (__U, _mm_and_si128 (__A, __B), |
||
| 2524 | _mm_setzero_si128()); |
||
| 2525 | } |
||
| 2526 | |||
| 2527 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 |
||
| 2528 | _mm256_testn_epi8_mask (__m256i __A, __m256i __B) |
||
| 2529 | { |
||
| 2530 | return _mm256_cmpeq_epi8_mask (_mm256_and_si256 (__A, __B), |
||
| 2531 | _mm256_setzero_si256()); |
||
| 2532 | } |
||
| 2533 | |||
| 2534 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 |
||
| 2535 | _mm256_mask_testn_epi8_mask (__mmask32 __U, __m256i __A, __m256i __B) |
||
| 2536 | { |
||
| 2537 | return _mm256_mask_cmpeq_epi8_mask (__U, _mm256_and_si256 (__A, __B), |
||
| 2538 | _mm256_setzero_si256()); |
||
| 2539 | } |
||
| 2540 | |||
| 2541 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 |
||
| 2542 | _mm_testn_epi16_mask (__m128i __A, __m128i __B) |
||
| 2543 | { |
||
| 2544 | return _mm_cmpeq_epi16_mask (_mm_and_si128 (__A, __B), _mm_setzero_si128()); |
||
| 2545 | } |
||
| 2546 | |||
| 2547 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 |
||
| 2548 | _mm_mask_testn_epi16_mask (__mmask8 __U, __m128i __A, __m128i __B) |
||
| 2549 | { |
||
| 2550 | return _mm_mask_cmpeq_epi16_mask (__U, _mm_and_si128(__A, __B), _mm_setzero_si128()); |
||
| 2551 | } |
||
| 2552 | |||
| 2553 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 |
||
| 2554 | _mm256_testn_epi16_mask (__m256i __A, __m256i __B) |
||
| 2555 | { |
||
| 2556 | return _mm256_cmpeq_epi16_mask (_mm256_and_si256(__A, __B), |
||
| 2557 | _mm256_setzero_si256()); |
||
| 2558 | } |
||
| 2559 | |||
| 2560 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 |
||
| 2561 | _mm256_mask_testn_epi16_mask (__mmask16 __U, __m256i __A, __m256i __B) |
||
| 2562 | { |
||
| 2563 | return _mm256_mask_cmpeq_epi16_mask (__U, _mm256_and_si256 (__A, __B), |
||
| 2564 | _mm256_setzero_si256()); |
||
| 2565 | } |
||
| 2566 | |||
| 2567 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS128 |
||
| 2568 | _mm_movepi8_mask (__m128i __A) |
||
| 2569 | { |
||
| 2570 | return (__mmask16) __builtin_ia32_cvtb2mask128 ((__v16qi) __A); |
||
| 2571 | } |
||
| 2572 | |||
| 2573 | static __inline__ __mmask32 __DEFAULT_FN_ATTRS256 |
||
| 2574 | _mm256_movepi8_mask (__m256i __A) |
||
| 2575 | { |
||
| 2576 | return (__mmask32) __builtin_ia32_cvtb2mask256 ((__v32qi) __A); |
||
| 2577 | } |
||
| 2578 | |||
| 2579 | static __inline__ __mmask8 __DEFAULT_FN_ATTRS128 |
||
| 2580 | _mm_movepi16_mask (__m128i __A) |
||
| 2581 | { |
||
| 2582 | return (__mmask8) __builtin_ia32_cvtw2mask128 ((__v8hi) __A); |
||
| 2583 | } |
||
| 2584 | |||
| 2585 | static __inline__ __mmask16 __DEFAULT_FN_ATTRS256 |
||
| 2586 | _mm256_movepi16_mask (__m256i __A) |
||
| 2587 | { |
||
| 2588 | return (__mmask16) __builtin_ia32_cvtw2mask256 ((__v16hi) __A); |
||
| 2589 | } |
||
| 2590 | |||
| 2591 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2592 | _mm_movm_epi8 (__mmask16 __A) |
||
| 2593 | { |
||
| 2594 | return (__m128i) __builtin_ia32_cvtmask2b128 (__A); |
||
| 2595 | } |
||
| 2596 | |||
| 2597 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2598 | _mm256_movm_epi8 (__mmask32 __A) |
||
| 2599 | { |
||
| 2600 | return (__m256i) __builtin_ia32_cvtmask2b256 (__A); |
||
| 2601 | } |
||
| 2602 | |||
| 2603 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2604 | _mm_movm_epi16 (__mmask8 __A) |
||
| 2605 | { |
||
| 2606 | return (__m128i) __builtin_ia32_cvtmask2w128 (__A); |
||
| 2607 | } |
||
| 2608 | |||
| 2609 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2610 | _mm256_movm_epi16 (__mmask16 __A) |
||
| 2611 | { |
||
| 2612 | return (__m256i) __builtin_ia32_cvtmask2w256 (__A); |
||
| 2613 | } |
||
| 2614 | |||
| 2615 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2616 | _mm_mask_broadcastb_epi8 (__m128i __O, __mmask16 __M, __m128i __A) |
||
| 2617 | { |
||
| 2618 | return (__m128i)__builtin_ia32_selectb_128(__M, |
||
| 2619 | (__v16qi) _mm_broadcastb_epi8(__A), |
||
| 2620 | (__v16qi) __O); |
||
| 2621 | } |
||
| 2622 | |||
| 2623 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2624 | _mm_maskz_broadcastb_epi8 (__mmask16 __M, __m128i __A) |
||
| 2625 | { |
||
| 2626 | return (__m128i)__builtin_ia32_selectb_128(__M, |
||
| 2627 | (__v16qi) _mm_broadcastb_epi8(__A), |
||
| 2628 | (__v16qi) _mm_setzero_si128()); |
||
| 2629 | } |
||
| 2630 | |||
| 2631 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2632 | _mm256_mask_broadcastb_epi8 (__m256i __O, __mmask32 __M, __m128i __A) |
||
| 2633 | { |
||
| 2634 | return (__m256i)__builtin_ia32_selectb_256(__M, |
||
| 2635 | (__v32qi) _mm256_broadcastb_epi8(__A), |
||
| 2636 | (__v32qi) __O); |
||
| 2637 | } |
||
| 2638 | |||
| 2639 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2640 | _mm256_maskz_broadcastb_epi8 (__mmask32 __M, __m128i __A) |
||
| 2641 | { |
||
| 2642 | return (__m256i)__builtin_ia32_selectb_256(__M, |
||
| 2643 | (__v32qi) _mm256_broadcastb_epi8(__A), |
||
| 2644 | (__v32qi) _mm256_setzero_si256()); |
||
| 2645 | } |
||
| 2646 | |||
| 2647 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2648 | _mm_mask_broadcastw_epi16 (__m128i __O, __mmask8 __M, __m128i __A) |
||
| 2649 | { |
||
| 2650 | return (__m128i)__builtin_ia32_selectw_128(__M, |
||
| 2651 | (__v8hi) _mm_broadcastw_epi16(__A), |
||
| 2652 | (__v8hi) __O); |
||
| 2653 | } |
||
| 2654 | |||
| 2655 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2656 | _mm_maskz_broadcastw_epi16 (__mmask8 __M, __m128i __A) |
||
| 2657 | { |
||
| 2658 | return (__m128i)__builtin_ia32_selectw_128(__M, |
||
| 2659 | (__v8hi) _mm_broadcastw_epi16(__A), |
||
| 2660 | (__v8hi) _mm_setzero_si128()); |
||
| 2661 | } |
||
| 2662 | |||
| 2663 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2664 | _mm256_mask_broadcastw_epi16 (__m256i __O, __mmask16 __M, __m128i __A) |
||
| 2665 | { |
||
| 2666 | return (__m256i)__builtin_ia32_selectw_256(__M, |
||
| 2667 | (__v16hi) _mm256_broadcastw_epi16(__A), |
||
| 2668 | (__v16hi) __O); |
||
| 2669 | } |
||
| 2670 | |||
| 2671 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2672 | _mm256_maskz_broadcastw_epi16 (__mmask16 __M, __m128i __A) |
||
| 2673 | { |
||
| 2674 | return (__m256i)__builtin_ia32_selectw_256(__M, |
||
| 2675 | (__v16hi) _mm256_broadcastw_epi16(__A), |
||
| 2676 | (__v16hi) _mm256_setzero_si256()); |
||
| 2677 | } |
||
| 2678 | |||
| 2679 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2680 | _mm256_mask_set1_epi16 (__m256i __O, __mmask16 __M, short __A) |
||
| 2681 | { |
||
| 2682 | return (__m256i) __builtin_ia32_selectw_256 (__M, |
||
| 2683 | (__v16hi) _mm256_set1_epi16(__A), |
||
| 2684 | (__v16hi) __O); |
||
| 2685 | } |
||
| 2686 | |||
| 2687 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2688 | _mm256_maskz_set1_epi16 (__mmask16 __M, short __A) |
||
| 2689 | { |
||
| 2690 | return (__m256i) __builtin_ia32_selectw_256(__M, |
||
| 2691 | (__v16hi)_mm256_set1_epi16(__A), |
||
| 2692 | (__v16hi) _mm256_setzero_si256()); |
||
| 2693 | } |
||
| 2694 | |||
| 2695 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2696 | _mm_mask_set1_epi16 (__m128i __O, __mmask8 __M, short __A) |
||
| 2697 | { |
||
| 2698 | return (__m128i) __builtin_ia32_selectw_128(__M, |
||
| 2699 | (__v8hi) _mm_set1_epi16(__A), |
||
| 2700 | (__v8hi) __O); |
||
| 2701 | } |
||
| 2702 | |||
| 2703 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2704 | _mm_maskz_set1_epi16 (__mmask8 __M, short __A) |
||
| 2705 | { |
||
| 2706 | return (__m128i) __builtin_ia32_selectw_128(__M, |
||
| 2707 | (__v8hi) _mm_set1_epi16(__A), |
||
| 2708 | (__v8hi) _mm_setzero_si128()); |
||
| 2709 | } |
||
| 2710 | |||
| 2711 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2712 | _mm_permutexvar_epi16 (__m128i __A, __m128i __B) |
||
| 2713 | { |
||
| 2714 | return (__m128i)__builtin_ia32_permvarhi128((__v8hi) __B, (__v8hi) __A); |
||
| 2715 | } |
||
| 2716 | |||
| 2717 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2718 | _mm_maskz_permutexvar_epi16 (__mmask8 __M, __m128i __A, __m128i __B) |
||
| 2719 | { |
||
| 2720 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 2721 | (__v8hi)_mm_permutexvar_epi16(__A, __B), |
||
| 2722 | (__v8hi) _mm_setzero_si128()); |
||
| 2723 | } |
||
| 2724 | |||
| 2725 | static __inline__ __m128i __DEFAULT_FN_ATTRS128 |
||
| 2726 | _mm_mask_permutexvar_epi16 (__m128i __W, __mmask8 __M, __m128i __A, |
||
| 2727 | __m128i __B) |
||
| 2728 | { |
||
| 2729 | return (__m128i)__builtin_ia32_selectw_128((__mmask8)__M, |
||
| 2730 | (__v8hi)_mm_permutexvar_epi16(__A, __B), |
||
| 2731 | (__v8hi)__W); |
||
| 2732 | } |
||
| 2733 | |||
| 2734 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2735 | _mm256_permutexvar_epi16 (__m256i __A, __m256i __B) |
||
| 2736 | { |
||
| 2737 | return (__m256i)__builtin_ia32_permvarhi256((__v16hi) __B, (__v16hi) __A); |
||
| 2738 | } |
||
| 2739 | |||
| 2740 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2741 | _mm256_maskz_permutexvar_epi16 (__mmask16 __M, __m256i __A, |
||
| 2742 | __m256i __B) |
||
| 2743 | { |
||
| 2744 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 2745 | (__v16hi)_mm256_permutexvar_epi16(__A, __B), |
||
| 2746 | (__v16hi)_mm256_setzero_si256()); |
||
| 2747 | } |
||
| 2748 | |||
| 2749 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 2750 | _mm256_mask_permutexvar_epi16 (__m256i __W, __mmask16 __M, __m256i __A, |
||
| 2751 | __m256i __B) |
||
| 2752 | { |
||
| 2753 | return (__m256i)__builtin_ia32_selectw_256((__mmask16)__M, |
||
| 2754 | (__v16hi)_mm256_permutexvar_epi16(__A, __B), |
||
| 2755 | (__v16hi)__W); |
||
| 2756 | } |
||
| 2757 | |||
| 2758 | #define _mm_mask_alignr_epi8(W, U, A, B, N) \ |
||
| 2759 | ((__m128i)__builtin_ia32_selectb_128((__mmask16)(U), \ |
||
| 2760 | (__v16qi)_mm_alignr_epi8((A), (B), (int)(N)), \ |
||
| 2761 | (__v16qi)(__m128i)(W))) |
||
| 2762 | |||
| 2763 | #define _mm_maskz_alignr_epi8(U, A, B, N) \ |
||
| 2764 | ((__m128i)__builtin_ia32_selectb_128((__mmask16)(U), \ |
||
| 2765 | (__v16qi)_mm_alignr_epi8((A), (B), (int)(N)), \ |
||
| 2766 | (__v16qi)_mm_setzero_si128())) |
||
| 2767 | |||
| 2768 | #define _mm256_mask_alignr_epi8(W, U, A, B, N) \ |
||
| 2769 | ((__m256i)__builtin_ia32_selectb_256((__mmask32)(U), \ |
||
| 2770 | (__v32qi)_mm256_alignr_epi8((A), (B), (int)(N)), \ |
||
| 2771 | (__v32qi)(__m256i)(W))) |
||
| 2772 | |||
| 2773 | #define _mm256_maskz_alignr_epi8(U, A, B, N) \ |
||
| 2774 | ((__m256i)__builtin_ia32_selectb_256((__mmask32)(U), \ |
||
| 2775 | (__v32qi)_mm256_alignr_epi8((A), (B), (int)(N)), \ |
||
| 2776 | (__v32qi)_mm256_setzero_si256())) |
||
| 2777 | |||
| 2778 | #define _mm_dbsad_epu8(A, B, imm) \ |
||
| 2779 | ((__m128i)__builtin_ia32_dbpsadbw128((__v16qi)(__m128i)(A), \ |
||
| 2780 | (__v16qi)(__m128i)(B), (int)(imm))) |
||
| 2781 | |||
| 2782 | #define _mm_mask_dbsad_epu8(W, U, A, B, imm) \ |
||
| 2783 | ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \ |
||
| 2784 | (__v8hi)_mm_dbsad_epu8((A), (B), (imm)), \ |
||
| 2785 | (__v8hi)(__m128i)(W))) |
||
| 2786 | |||
| 2787 | #define _mm_maskz_dbsad_epu8(U, A, B, imm) \ |
||
| 2788 | ((__m128i)__builtin_ia32_selectw_128((__mmask8)(U), \ |
||
| 2789 | (__v8hi)_mm_dbsad_epu8((A), (B), (imm)), \ |
||
| 2790 | (__v8hi)_mm_setzero_si128())) |
||
| 2791 | |||
| 2792 | #define _mm256_dbsad_epu8(A, B, imm) \ |
||
| 2793 | ((__m256i)__builtin_ia32_dbpsadbw256((__v32qi)(__m256i)(A), \ |
||
| 2794 | (__v32qi)(__m256i)(B), (int)(imm))) |
||
| 2795 | |||
| 2796 | #define _mm256_mask_dbsad_epu8(W, U, A, B, imm) \ |
||
| 2797 | ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \ |
||
| 2798 | (__v16hi)_mm256_dbsad_epu8((A), (B), (imm)), \ |
||
| 2799 | (__v16hi)(__m256i)(W))) |
||
| 2800 | |||
| 2801 | #define _mm256_maskz_dbsad_epu8(U, A, B, imm) \ |
||
| 2802 | ((__m256i)__builtin_ia32_selectw_256((__mmask16)(U), \ |
||
| 2803 | (__v16hi)_mm256_dbsad_epu8((A), (B), (imm)), \ |
||
| 2804 | (__v16hi)_mm256_setzero_si256())) |
||
| 2805 | |||
| 2806 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2807 | _mm_reduce_add_epi16(__m128i __W) { |
||
| 2808 | return __builtin_reduce_add((__v8hi)__W); |
||
| 2809 | } |
||
| 2810 | |||
| 2811 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2812 | _mm_reduce_mul_epi16(__m128i __W) { |
||
| 2813 | return __builtin_reduce_mul((__v8hi)__W); |
||
| 2814 | } |
||
| 2815 | |||
| 2816 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2817 | _mm_reduce_and_epi16(__m128i __W) { |
||
| 2818 | return __builtin_reduce_and((__v8hi)__W); |
||
| 2819 | } |
||
| 2820 | |||
| 2821 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2822 | _mm_reduce_or_epi16(__m128i __W) { |
||
| 2823 | return __builtin_reduce_or((__v8hi)__W); |
||
| 2824 | } |
||
| 2825 | |||
| 2826 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2827 | _mm_mask_reduce_add_epi16( __mmask8 __M, __m128i __W) { |
||
| 2828 | __W = _mm_maskz_mov_epi16(__M, __W); |
||
| 2829 | return __builtin_reduce_add((__v8hi)__W); |
||
| 2830 | } |
||
| 2831 | |||
| 2832 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2833 | _mm_mask_reduce_mul_epi16( __mmask8 __M, __m128i __W) { |
||
| 2834 | __W = _mm_mask_mov_epi16(_mm_set1_epi16(1), __M, __W); |
||
| 2835 | return __builtin_reduce_mul((__v8hi)__W); |
||
| 2836 | } |
||
| 2837 | |||
| 2838 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2839 | _mm_mask_reduce_and_epi16( __mmask8 __M, __m128i __W) { |
||
| 2840 | __W = _mm_mask_mov_epi16(_mm_set1_epi16(-1), __M, __W); |
||
| 2841 | return __builtin_reduce_and((__v8hi)__W); |
||
| 2842 | } |
||
| 2843 | |||
| 2844 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2845 | _mm_mask_reduce_or_epi16(__mmask8 __M, __m128i __W) { |
||
| 2846 | __W = _mm_maskz_mov_epi16(__M, __W); |
||
| 2847 | return __builtin_reduce_or((__v8hi)__W); |
||
| 2848 | } |
||
| 2849 | |||
| 2850 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2851 | _mm_reduce_max_epi16(__m128i __V) { |
||
| 2852 | return __builtin_reduce_max((__v8hi)__V); |
||
| 2853 | } |
||
| 2854 | |||
| 2855 | static __inline__ unsigned short __DEFAULT_FN_ATTRS128 |
||
| 2856 | _mm_reduce_max_epu16(__m128i __V) { |
||
| 2857 | return __builtin_reduce_max((__v8hu)__V); |
||
| 2858 | } |
||
| 2859 | |||
| 2860 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2861 | _mm_reduce_min_epi16(__m128i __V) { |
||
| 2862 | return __builtin_reduce_min((__v8hi)__V); |
||
| 2863 | } |
||
| 2864 | |||
| 2865 | static __inline__ unsigned short __DEFAULT_FN_ATTRS128 |
||
| 2866 | _mm_reduce_min_epu16(__m128i __V) { |
||
| 2867 | return __builtin_reduce_min((__v8hu)__V); |
||
| 2868 | } |
||
| 2869 | |||
| 2870 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2871 | _mm_mask_reduce_max_epi16(__mmask16 __M, __m128i __V) { |
||
| 2872 | __V = _mm_mask_mov_epi16(_mm_set1_epi16(-32767-1), __M, __V); |
||
| 2873 | return __builtin_reduce_max((__v8hi)__V); |
||
| 2874 | } |
||
| 2875 | |||
| 2876 | static __inline__ unsigned short __DEFAULT_FN_ATTRS128 |
||
| 2877 | _mm_mask_reduce_max_epu16(__mmask16 __M, __m128i __V) { |
||
| 2878 | __V = _mm_maskz_mov_epi16(__M, __V); |
||
| 2879 | return __builtin_reduce_max((__v8hu)__V); |
||
| 2880 | } |
||
| 2881 | |||
| 2882 | static __inline__ short __DEFAULT_FN_ATTRS128 |
||
| 2883 | _mm_mask_reduce_min_epi16(__mmask16 __M, __m128i __V) { |
||
| 2884 | __V = _mm_mask_mov_epi16(_mm_set1_epi16(32767), __M, __V); |
||
| 2885 | return __builtin_reduce_min((__v8hi)__V); |
||
| 2886 | } |
||
| 2887 | |||
| 2888 | static __inline__ unsigned short __DEFAULT_FN_ATTRS128 |
||
| 2889 | _mm_mask_reduce_min_epu16(__mmask16 __M, __m128i __V) { |
||
| 2890 | __V = _mm_mask_mov_epi16(_mm_set1_epi16(-1), __M, __V); |
||
| 2891 | return __builtin_reduce_min((__v8hu)__V); |
||
| 2892 | } |
||
| 2893 | |||
| 2894 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2895 | _mm256_reduce_add_epi16(__m256i __W) { |
||
| 2896 | return __builtin_reduce_add((__v16hi)__W); |
||
| 2897 | } |
||
| 2898 | |||
| 2899 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2900 | _mm256_reduce_mul_epi16(__m256i __W) { |
||
| 2901 | return __builtin_reduce_mul((__v16hi)__W); |
||
| 2902 | } |
||
| 2903 | |||
| 2904 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2905 | _mm256_reduce_and_epi16(__m256i __W) { |
||
| 2906 | return __builtin_reduce_and((__v16hi)__W); |
||
| 2907 | } |
||
| 2908 | |||
| 2909 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2910 | _mm256_reduce_or_epi16(__m256i __W) { |
||
| 2911 | return __builtin_reduce_or((__v16hi)__W); |
||
| 2912 | } |
||
| 2913 | |||
| 2914 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2915 | _mm256_mask_reduce_add_epi16( __mmask16 __M, __m256i __W) { |
||
| 2916 | __W = _mm256_maskz_mov_epi16(__M, __W); |
||
| 2917 | return __builtin_reduce_add((__v16hi)__W); |
||
| 2918 | } |
||
| 2919 | |||
| 2920 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2921 | _mm256_mask_reduce_mul_epi16( __mmask16 __M, __m256i __W) { |
||
| 2922 | __W = _mm256_mask_mov_epi16(_mm256_set1_epi16(1), __M, __W); |
||
| 2923 | return __builtin_reduce_mul((__v16hi)__W); |
||
| 2924 | } |
||
| 2925 | |||
| 2926 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2927 | _mm256_mask_reduce_and_epi16( __mmask16 __M, __m256i __W) { |
||
| 2928 | __W = _mm256_mask_mov_epi16(_mm256_set1_epi16(-1), __M, __W); |
||
| 2929 | return __builtin_reduce_and((__v16hi)__W); |
||
| 2930 | } |
||
| 2931 | |||
| 2932 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2933 | _mm256_mask_reduce_or_epi16(__mmask16 __M, __m256i __W) { |
||
| 2934 | __W = _mm256_maskz_mov_epi16(__M, __W); |
||
| 2935 | return __builtin_reduce_or((__v16hi)__W); |
||
| 2936 | } |
||
| 2937 | |||
| 2938 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2939 | _mm256_reduce_max_epi16(__m256i __V) { |
||
| 2940 | return __builtin_reduce_max((__v16hi)__V); |
||
| 2941 | } |
||
| 2942 | |||
| 2943 | static __inline__ unsigned short __DEFAULT_FN_ATTRS256 |
||
| 2944 | _mm256_reduce_max_epu16(__m256i __V) { |
||
| 2945 | return __builtin_reduce_max((__v16hu)__V); |
||
| 2946 | } |
||
| 2947 | |||
| 2948 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2949 | _mm256_reduce_min_epi16(__m256i __V) { |
||
| 2950 | return __builtin_reduce_min((__v16hi)__V); |
||
| 2951 | } |
||
| 2952 | |||
| 2953 | static __inline__ unsigned short __DEFAULT_FN_ATTRS256 |
||
| 2954 | _mm256_reduce_min_epu16(__m256i __V) { |
||
| 2955 | return __builtin_reduce_min((__v16hu)__V); |
||
| 2956 | } |
||
| 2957 | |||
| 2958 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2959 | _mm256_mask_reduce_max_epi16(__mmask16 __M, __m256i __V) { |
||
| 2960 | __V = _mm256_mask_mov_epi16(_mm256_set1_epi16(-32767-1), __M, __V); |
||
| 2961 | return __builtin_reduce_max((__v16hi)__V); |
||
| 2962 | } |
||
| 2963 | |||
| 2964 | static __inline__ unsigned short __DEFAULT_FN_ATTRS256 |
||
| 2965 | _mm256_mask_reduce_max_epu16(__mmask16 __M, __m256i __V) { |
||
| 2966 | __V = _mm256_maskz_mov_epi16(__M, __V); |
||
| 2967 | return __builtin_reduce_max((__v16hu)__V); |
||
| 2968 | } |
||
| 2969 | |||
| 2970 | static __inline__ short __DEFAULT_FN_ATTRS256 |
||
| 2971 | _mm256_mask_reduce_min_epi16(__mmask16 __M, __m256i __V) { |
||
| 2972 | __V = _mm256_mask_mov_epi16(_mm256_set1_epi16(32767), __M, __V); |
||
| 2973 | return __builtin_reduce_min((__v16hi)__V); |
||
| 2974 | } |
||
| 2975 | |||
| 2976 | static __inline__ unsigned short __DEFAULT_FN_ATTRS256 |
||
| 2977 | _mm256_mask_reduce_min_epu16(__mmask16 __M, __m256i __V) { |
||
| 2978 | __V = _mm256_mask_mov_epi16(_mm256_set1_epi16(-1), __M, __V); |
||
| 2979 | return __builtin_reduce_min((__v16hu)__V); |
||
| 2980 | } |
||
| 2981 | |||
| 2982 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 2983 | _mm_reduce_add_epi8(__m128i __W) { |
||
| 2984 | return __builtin_reduce_add((__v16qs)__W); |
||
| 2985 | } |
||
| 2986 | |||
| 2987 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 2988 | _mm_reduce_mul_epi8(__m128i __W) { |
||
| 2989 | return __builtin_reduce_mul((__v16qs)__W); |
||
| 2990 | } |
||
| 2991 | |||
| 2992 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 2993 | _mm_reduce_and_epi8(__m128i __W) { |
||
| 2994 | return __builtin_reduce_and((__v16qs)__W); |
||
| 2995 | } |
||
| 2996 | |||
| 2997 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 2998 | _mm_reduce_or_epi8(__m128i __W) { |
||
| 2999 | return __builtin_reduce_or((__v16qs)__W); |
||
| 3000 | } |
||
| 3001 | |||
| 3002 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3003 | _mm_mask_reduce_add_epi8(__mmask16 __M, __m128i __W) { |
||
| 3004 | __W = _mm_maskz_mov_epi8(__M, __W); |
||
| 3005 | return __builtin_reduce_add((__v16qs)__W); |
||
| 3006 | } |
||
| 3007 | |||
| 3008 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3009 | _mm_mask_reduce_mul_epi8(__mmask16 __M, __m128i __W) { |
||
| 3010 | __W = _mm_mask_mov_epi8(_mm_set1_epi8(1), __M, __W); |
||
| 3011 | return __builtin_reduce_mul((__v16qs)__W); |
||
| 3012 | } |
||
| 3013 | |||
| 3014 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3015 | _mm_mask_reduce_and_epi8(__mmask16 __M, __m128i __W) { |
||
| 3016 | __W = _mm_mask_mov_epi8(_mm_set1_epi8(-1), __M, __W); |
||
| 3017 | return __builtin_reduce_and((__v16qs)__W); |
||
| 3018 | } |
||
| 3019 | |||
| 3020 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3021 | _mm_mask_reduce_or_epi8(__mmask16 __M, __m128i __W) { |
||
| 3022 | __W = _mm_maskz_mov_epi8(__M, __W); |
||
| 3023 | return __builtin_reduce_or((__v16qs)__W); |
||
| 3024 | } |
||
| 3025 | |||
| 3026 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3027 | _mm_reduce_max_epi8(__m128i __V) { |
||
| 3028 | return __builtin_reduce_max((__v16qs)__V); |
||
| 3029 | } |
||
| 3030 | |||
| 3031 | static __inline__ unsigned char __DEFAULT_FN_ATTRS128 |
||
| 3032 | _mm_reduce_max_epu8(__m128i __V) { |
||
| 3033 | return __builtin_reduce_max((__v16qu)__V); |
||
| 3034 | } |
||
| 3035 | |||
| 3036 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3037 | _mm_reduce_min_epi8(__m128i __V) { |
||
| 3038 | return __builtin_reduce_min((__v16qs)__V); |
||
| 3039 | } |
||
| 3040 | |||
| 3041 | static __inline__ unsigned char __DEFAULT_FN_ATTRS128 |
||
| 3042 | _mm_reduce_min_epu8(__m128i __V) { |
||
| 3043 | return __builtin_reduce_min((__v16qu)__V); |
||
| 3044 | } |
||
| 3045 | |||
| 3046 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3047 | _mm_mask_reduce_max_epi8(__mmask16 __M, __m128i __V) { |
||
| 3048 | __V = _mm_mask_mov_epi8(_mm_set1_epi8(-127-1), __M, __V); |
||
| 3049 | return __builtin_reduce_max((__v16qs)__V); |
||
| 3050 | } |
||
| 3051 | |||
| 3052 | static __inline__ unsigned char __DEFAULT_FN_ATTRS128 |
||
| 3053 | _mm_mask_reduce_max_epu8(__mmask16 __M, __m128i __V) { |
||
| 3054 | __V = _mm_maskz_mov_epi8(__M, __V); |
||
| 3055 | return __builtin_reduce_max((__v16qu)__V); |
||
| 3056 | } |
||
| 3057 | |||
| 3058 | static __inline__ signed char __DEFAULT_FN_ATTRS128 |
||
| 3059 | _mm_mask_reduce_min_epi8(__mmask16 __M, __m128i __V) { |
||
| 3060 | __V = _mm_mask_mov_epi8(_mm_set1_epi8(127), __M, __V); |
||
| 3061 | return __builtin_reduce_min((__v16qs)__V); |
||
| 3062 | } |
||
| 3063 | |||
| 3064 | static __inline__ unsigned char __DEFAULT_FN_ATTRS128 |
||
| 3065 | _mm_mask_reduce_min_epu8(__mmask16 __M, __m128i __V) { |
||
| 3066 | __V = _mm_mask_mov_epi8(_mm_set1_epi8(-1), __M, __V); |
||
| 3067 | return __builtin_reduce_min((__v16qu)__V); |
||
| 3068 | } |
||
| 3069 | |||
| 3070 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3071 | _mm256_reduce_add_epi8(__m256i __W) { |
||
| 3072 | return __builtin_reduce_add((__v32qs)__W); |
||
| 3073 | } |
||
| 3074 | |||
| 3075 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3076 | _mm256_reduce_mul_epi8(__m256i __W) { |
||
| 3077 | return __builtin_reduce_mul((__v32qs)__W); |
||
| 3078 | } |
||
| 3079 | |||
| 3080 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3081 | _mm256_reduce_and_epi8(__m256i __W) { |
||
| 3082 | return __builtin_reduce_and((__v32qs)__W); |
||
| 3083 | } |
||
| 3084 | |||
| 3085 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3086 | _mm256_reduce_or_epi8(__m256i __W) { |
||
| 3087 | return __builtin_reduce_or((__v32qs)__W); |
||
| 3088 | } |
||
| 3089 | |||
| 3090 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3091 | _mm256_mask_reduce_add_epi8(__mmask32 __M, __m256i __W) { |
||
| 3092 | __W = _mm256_maskz_mov_epi8(__M, __W); |
||
| 3093 | return __builtin_reduce_add((__v32qs)__W); |
||
| 3094 | } |
||
| 3095 | |||
| 3096 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3097 | _mm256_mask_reduce_mul_epi8(__mmask32 __M, __m256i __W) { |
||
| 3098 | __W = _mm256_mask_mov_epi8(_mm256_set1_epi8(1), __M, __W); |
||
| 3099 | return __builtin_reduce_mul((__v32qs)__W); |
||
| 3100 | } |
||
| 3101 | |||
| 3102 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3103 | _mm256_mask_reduce_and_epi8(__mmask32 __M, __m256i __W) { |
||
| 3104 | __W = _mm256_mask_mov_epi8(_mm256_set1_epi8(-1), __M, __W); |
||
| 3105 | return __builtin_reduce_and((__v32qs)__W); |
||
| 3106 | } |
||
| 3107 | |||
| 3108 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3109 | _mm256_mask_reduce_or_epi8(__mmask32 __M, __m256i __W) { |
||
| 3110 | __W = _mm256_maskz_mov_epi8(__M, __W); |
||
| 3111 | return __builtin_reduce_or((__v32qs)__W); |
||
| 3112 | } |
||
| 3113 | |||
| 3114 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3115 | _mm256_reduce_max_epi8(__m256i __V) { |
||
| 3116 | return __builtin_reduce_max((__v32qs)__V); |
||
| 3117 | } |
||
| 3118 | |||
| 3119 | static __inline__ unsigned char __DEFAULT_FN_ATTRS256 |
||
| 3120 | _mm256_reduce_max_epu8(__m256i __V) { |
||
| 3121 | return __builtin_reduce_max((__v32qu)__V); |
||
| 3122 | } |
||
| 3123 | |||
| 3124 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3125 | _mm256_reduce_min_epi8(__m256i __V) { |
||
| 3126 | return __builtin_reduce_min((__v32qs)__V); |
||
| 3127 | } |
||
| 3128 | |||
| 3129 | static __inline__ unsigned char __DEFAULT_FN_ATTRS256 |
||
| 3130 | _mm256_reduce_min_epu8(__m256i __V) { |
||
| 3131 | return __builtin_reduce_min((__v32qu)__V); |
||
| 3132 | } |
||
| 3133 | |||
| 3134 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3135 | _mm256_mask_reduce_max_epi8(__mmask32 __M, __m256i __V) { |
||
| 3136 | __V = _mm256_mask_mov_epi8(_mm256_set1_epi8(-127-1), __M, __V); |
||
| 3137 | return __builtin_reduce_max((__v32qs)__V); |
||
| 3138 | } |
||
| 3139 | |||
| 3140 | static __inline__ unsigned char __DEFAULT_FN_ATTRS256 |
||
| 3141 | _mm256_mask_reduce_max_epu8(__mmask32 __M, __m256i __V) { |
||
| 3142 | __V = _mm256_maskz_mov_epi8(__M, __V); |
||
| 3143 | return __builtin_reduce_max((__v32qu)__V); |
||
| 3144 | } |
||
| 3145 | |||
| 3146 | static __inline__ signed char __DEFAULT_FN_ATTRS256 |
||
| 3147 | _mm256_mask_reduce_min_epi8(__mmask32 __M, __m256i __V) { |
||
| 3148 | __V = _mm256_mask_mov_epi8(_mm256_set1_epi8(127), __M, __V); |
||
| 3149 | return __builtin_reduce_min((__v32qs)__V); |
||
| 3150 | } |
||
| 3151 | |||
| 3152 | static __inline__ unsigned char __DEFAULT_FN_ATTRS256 |
||
| 3153 | _mm256_mask_reduce_min_epu8(__mmask32 __M, __m256i __V) { |
||
| 3154 | __V = _mm256_mask_mov_epi8(_mm256_set1_epi8(-1), __M, __V); |
||
| 3155 | return __builtin_reduce_min((__v32qu)__V); |
||
| 3156 | } |
||
| 3157 | |||
| 3158 | #undef __DEFAULT_FN_ATTRS128 |
||
| 3159 | #undef __DEFAULT_FN_ATTRS256 |
||
| 3160 | |||
| 3161 | #endif /* __AVX512VLBWINTRIN_H */ |