Stephen Hines | ee4ca28 | 2014-12-02 17:05:12 -0800 | [diff] [blame] | 1 | /*===---- avx512vlbwintrin.h - AVX512VL and AVX512BW intrinsics ----------=== |
| 2 | * |
| 3 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 4 | * of this software and associated documentation files (the "Software"), to deal |
| 5 | * in the Software without restriction, including without limitation the rights |
| 6 | * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell |
| 7 | * copies of the Software, and to permit persons to whom the Software is |
| 8 | * furnished to do so, subject to the following conditions: |
| 9 | * |
| 10 | * The above copyright notice and this permission notice shall be included in |
| 11 | * all copies or substantial portions of the Software. |
| 12 | * |
| 13 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 14 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 15 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 16 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 17 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 18 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN |
| 19 | * THE SOFTWARE. |
| 20 | * |
| 21 | *===-----------------------------------------------------------------------=== |
| 22 | */ |
| 23 | |
| 24 | #ifndef __IMMINTRIN_H |
| 25 | #error "Never use <avx512vlbwintrin.h> directly; include <immintrin.h> instead." |
| 26 | #endif |
| 27 | |
| 28 | #ifndef __AVX512VLBWINTRIN_H |
| 29 | #define __AVX512VLBWINTRIN_H |
| 30 | |
| 31 | /* Integer compare */ |
| 32 | |
| 33 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 34 | _mm_cmpeq_epi8_mask(__m128i __a, __m128i __b) { |
| 35 | return (__mmask16)__builtin_ia32_pcmpeqb128_mask((__v16qi)__a, (__v16qi)__b, |
| 36 | (__mmask16)-1); |
| 37 | } |
| 38 | |
| 39 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 40 | _mm_mask_cmpeq_epi8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 41 | return (__mmask16)__builtin_ia32_pcmpeqb128_mask((__v16qi)__a, (__v16qi)__b, |
| 42 | __u); |
| 43 | } |
| 44 | |
Stephen Hines | b4d9c8b | 2015-03-30 16:04:04 -0700 | [diff] [blame] | 45 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 46 | _mm_cmpeq_epu8_mask(__m128i __a, __m128i __b) { |
| 47 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 0, |
| 48 | (__mmask16)-1); |
| 49 | } |
| 50 | |
| 51 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 52 | _mm_mask_cmpeq_epu8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 53 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 0, |
| 54 | __u); |
| 55 | } |
Stephen Hines | ee4ca28 | 2014-12-02 17:05:12 -0800 | [diff] [blame] | 56 | |
| 57 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 58 | _mm256_cmpeq_epi8_mask(__m256i __a, __m256i __b) { |
| 59 | return (__mmask32)__builtin_ia32_pcmpeqb256_mask((__v32qi)__a, (__v32qi)__b, |
| 60 | (__mmask32)-1); |
| 61 | } |
| 62 | |
| 63 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 64 | _mm256_mask_cmpeq_epi8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 65 | return (__mmask32)__builtin_ia32_pcmpeqb256_mask((__v32qi)__a, (__v32qi)__b, |
| 66 | __u); |
| 67 | } |
| 68 | |
Stephen Hines | b4d9c8b | 2015-03-30 16:04:04 -0700 | [diff] [blame] | 69 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 70 | _mm256_cmpeq_epu8_mask(__m256i __a, __m256i __b) { |
| 71 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 0, |
| 72 | (__mmask32)-1); |
| 73 | } |
| 74 | |
| 75 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 76 | _mm256_mask_cmpeq_epu8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 77 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 0, |
| 78 | __u); |
| 79 | } |
| 80 | |
Stephen Hines | ee4ca28 | 2014-12-02 17:05:12 -0800 | [diff] [blame] | 81 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 82 | _mm_cmpeq_epi16_mask(__m128i __a, __m128i __b) { |
| 83 | return (__mmask8)__builtin_ia32_pcmpeqw128_mask((__v8hi)__a, (__v8hi)__b, |
| 84 | (__mmask8)-1); |
| 85 | } |
| 86 | |
| 87 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 88 | _mm_mask_cmpeq_epi16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 89 | return (__mmask8)__builtin_ia32_pcmpeqw128_mask((__v8hi)__a, (__v8hi)__b, |
| 90 | __u); |
| 91 | } |
| 92 | |
Stephen Hines | b4d9c8b | 2015-03-30 16:04:04 -0700 | [diff] [blame] | 93 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 94 | _mm_cmpeq_epu16_mask(__m128i __a, __m128i __b) { |
| 95 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 0, |
| 96 | (__mmask8)-1); |
| 97 | } |
| 98 | |
| 99 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 100 | _mm_mask_cmpeq_epu16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 101 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 0, |
| 102 | __u); |
| 103 | } |
Stephen Hines | ee4ca28 | 2014-12-02 17:05:12 -0800 | [diff] [blame] | 104 | |
| 105 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 106 | _mm256_cmpeq_epi16_mask(__m256i __a, __m256i __b) { |
| 107 | return (__mmask16)__builtin_ia32_pcmpeqw256_mask((__v16hi)__a, (__v16hi)__b, |
| 108 | (__mmask16)-1); |
| 109 | } |
| 110 | |
| 111 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 112 | _mm256_mask_cmpeq_epi16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 113 | return (__mmask16)__builtin_ia32_pcmpeqw256_mask((__v16hi)__a, (__v16hi)__b, |
| 114 | __u); |
| 115 | } |
| 116 | |
Stephen Hines | b4d9c8b | 2015-03-30 16:04:04 -0700 | [diff] [blame] | 117 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 118 | _mm256_cmpeq_epu16_mask(__m256i __a, __m256i __b) { |
| 119 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 0, |
| 120 | (__mmask16)-1); |
| 121 | } |
| 122 | |
| 123 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 124 | _mm256_mask_cmpeq_epu16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 125 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 0, |
| 126 | __u); |
| 127 | } |
| 128 | |
| 129 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 130 | _mm_cmpge_epi8_mask(__m128i __a, __m128i __b) { |
| 131 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 5, |
| 132 | (__mmask16)-1); |
| 133 | } |
| 134 | |
| 135 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 136 | _mm_mask_cmpge_epi8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 137 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 5, |
| 138 | __u); |
| 139 | } |
| 140 | |
| 141 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 142 | _mm_cmpge_epu8_mask(__m128i __a, __m128i __b) { |
| 143 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 5, |
| 144 | (__mmask16)-1); |
| 145 | } |
| 146 | |
| 147 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 148 | _mm_mask_cmpge_epu8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 149 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 5, |
| 150 | __u); |
| 151 | } |
| 152 | |
| 153 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 154 | _mm256_cmpge_epi8_mask(__m256i __a, __m256i __b) { |
| 155 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 5, |
| 156 | (__mmask32)-1); |
| 157 | } |
| 158 | |
| 159 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 160 | _mm256_mask_cmpge_epi8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 161 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 5, |
| 162 | __u); |
| 163 | } |
| 164 | |
| 165 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 166 | _mm256_cmpge_epu8_mask(__m256i __a, __m256i __b) { |
| 167 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 5, |
| 168 | (__mmask32)-1); |
| 169 | } |
| 170 | |
| 171 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 172 | _mm256_mask_cmpge_epu8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 173 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 5, |
| 174 | __u); |
| 175 | } |
| 176 | |
| 177 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 178 | _mm_cmpge_epi16_mask(__m128i __a, __m128i __b) { |
| 179 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 5, |
| 180 | (__mmask8)-1); |
| 181 | } |
| 182 | |
| 183 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 184 | _mm_mask_cmpge_epi16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 185 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 5, |
| 186 | __u); |
| 187 | } |
| 188 | |
| 189 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 190 | _mm_cmpge_epu16_mask(__m128i __a, __m128i __b) { |
| 191 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 5, |
| 192 | (__mmask8)-1); |
| 193 | } |
| 194 | |
| 195 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 196 | _mm_mask_cmpge_epu16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 197 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 5, |
| 198 | __u); |
| 199 | } |
| 200 | |
| 201 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 202 | _mm256_cmpge_epi16_mask(__m256i __a, __m256i __b) { |
| 203 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 5, |
| 204 | (__mmask16)-1); |
| 205 | } |
| 206 | |
| 207 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 208 | _mm256_mask_cmpge_epi16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 209 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 5, |
| 210 | __u); |
| 211 | } |
| 212 | |
| 213 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 214 | _mm256_cmpge_epu16_mask(__m256i __a, __m256i __b) { |
| 215 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 5, |
| 216 | (__mmask16)-1); |
| 217 | } |
| 218 | |
| 219 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 220 | _mm256_mask_cmpge_epu16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 221 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 5, |
| 222 | __u); |
| 223 | } |
| 224 | |
| 225 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 226 | _mm_cmpgt_epi8_mask(__m128i __a, __m128i __b) { |
| 227 | return (__mmask16)__builtin_ia32_pcmpgtb128_mask((__v16qi)__a, (__v16qi)__b, |
| 228 | (__mmask16)-1); |
| 229 | } |
| 230 | |
| 231 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 232 | _mm_mask_cmpgt_epi8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 233 | return (__mmask16)__builtin_ia32_pcmpgtb128_mask((__v16qi)__a, (__v16qi)__b, |
| 234 | __u); |
| 235 | } |
| 236 | |
| 237 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 238 | _mm_cmpgt_epu8_mask(__m128i __a, __m128i __b) { |
| 239 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 6, |
| 240 | (__mmask16)-1); |
| 241 | } |
| 242 | |
| 243 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 244 | _mm_mask_cmpgt_epu8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 245 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 6, |
| 246 | __u); |
| 247 | } |
| 248 | |
| 249 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 250 | _mm256_cmpgt_epi8_mask(__m256i __a, __m256i __b) { |
| 251 | return (__mmask32)__builtin_ia32_pcmpgtb256_mask((__v32qi)__a, (__v32qi)__b, |
| 252 | (__mmask32)-1); |
| 253 | } |
| 254 | |
| 255 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 256 | _mm256_mask_cmpgt_epi8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 257 | return (__mmask32)__builtin_ia32_pcmpgtb256_mask((__v32qi)__a, (__v32qi)__b, |
| 258 | __u); |
| 259 | } |
| 260 | |
| 261 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 262 | _mm256_cmpgt_epu8_mask(__m256i __a, __m256i __b) { |
| 263 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 6, |
| 264 | (__mmask32)-1); |
| 265 | } |
| 266 | |
| 267 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 268 | _mm256_mask_cmpgt_epu8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 269 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 6, |
| 270 | __u); |
| 271 | } |
| 272 | |
| 273 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 274 | _mm_cmpgt_epi16_mask(__m128i __a, __m128i __b) { |
| 275 | return (__mmask8)__builtin_ia32_pcmpgtw128_mask((__v8hi)__a, (__v8hi)__b, |
| 276 | (__mmask8)-1); |
| 277 | } |
| 278 | |
| 279 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 280 | _mm_mask_cmpgt_epi16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 281 | return (__mmask8)__builtin_ia32_pcmpgtw128_mask((__v8hi)__a, (__v8hi)__b, |
| 282 | __u); |
| 283 | } |
| 284 | |
| 285 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 286 | _mm_cmpgt_epu16_mask(__m128i __a, __m128i __b) { |
| 287 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 6, |
| 288 | (__mmask8)-1); |
| 289 | } |
| 290 | |
| 291 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 292 | _mm_mask_cmpgt_epu16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 293 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 6, |
| 294 | __u); |
| 295 | } |
| 296 | |
| 297 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 298 | _mm256_cmpgt_epi16_mask(__m256i __a, __m256i __b) { |
| 299 | return (__mmask16)__builtin_ia32_pcmpgtw256_mask((__v16hi)__a, (__v16hi)__b, |
| 300 | (__mmask16)-1); |
| 301 | } |
| 302 | |
| 303 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 304 | _mm256_mask_cmpgt_epi16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 305 | return (__mmask16)__builtin_ia32_pcmpgtw256_mask((__v16hi)__a, (__v16hi)__b, |
| 306 | __u); |
| 307 | } |
| 308 | |
| 309 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 310 | _mm256_cmpgt_epu16_mask(__m256i __a, __m256i __b) { |
| 311 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 6, |
| 312 | (__mmask16)-1); |
| 313 | } |
| 314 | |
| 315 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 316 | _mm256_mask_cmpgt_epu16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 317 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 6, |
| 318 | __u); |
| 319 | } |
| 320 | |
| 321 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 322 | _mm_cmple_epi8_mask(__m128i __a, __m128i __b) { |
| 323 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 2, |
| 324 | (__mmask16)-1); |
| 325 | } |
| 326 | |
| 327 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 328 | _mm_mask_cmple_epi8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 329 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 2, |
| 330 | __u); |
| 331 | } |
| 332 | |
| 333 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 334 | _mm_cmple_epu8_mask(__m128i __a, __m128i __b) { |
| 335 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 2, |
| 336 | (__mmask16)-1); |
| 337 | } |
| 338 | |
| 339 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 340 | _mm_mask_cmple_epu8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 341 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 2, |
| 342 | __u); |
| 343 | } |
| 344 | |
| 345 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 346 | _mm256_cmple_epi8_mask(__m256i __a, __m256i __b) { |
| 347 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 2, |
| 348 | (__mmask32)-1); |
| 349 | } |
| 350 | |
| 351 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 352 | _mm256_mask_cmple_epi8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 353 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 2, |
| 354 | __u); |
| 355 | } |
| 356 | |
| 357 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 358 | _mm256_cmple_epu8_mask(__m256i __a, __m256i __b) { |
| 359 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 2, |
| 360 | (__mmask32)-1); |
| 361 | } |
| 362 | |
| 363 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 364 | _mm256_mask_cmple_epu8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 365 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 2, |
| 366 | __u); |
| 367 | } |
| 368 | |
| 369 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 370 | _mm_cmple_epi16_mask(__m128i __a, __m128i __b) { |
| 371 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 2, |
| 372 | (__mmask8)-1); |
| 373 | } |
| 374 | |
| 375 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 376 | _mm_mask_cmple_epi16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 377 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 2, |
| 378 | __u); |
| 379 | } |
| 380 | |
| 381 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 382 | _mm_cmple_epu16_mask(__m128i __a, __m128i __b) { |
| 383 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 2, |
| 384 | (__mmask8)-1); |
| 385 | } |
| 386 | |
| 387 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 388 | _mm_mask_cmple_epu16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 389 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 2, |
| 390 | __u); |
| 391 | } |
| 392 | |
| 393 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 394 | _mm256_cmple_epi16_mask(__m256i __a, __m256i __b) { |
| 395 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 2, |
| 396 | (__mmask16)-1); |
| 397 | } |
| 398 | |
| 399 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 400 | _mm256_mask_cmple_epi16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 401 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 2, |
| 402 | __u); |
| 403 | } |
| 404 | |
| 405 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 406 | _mm256_cmple_epu16_mask(__m256i __a, __m256i __b) { |
| 407 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 2, |
| 408 | (__mmask16)-1); |
| 409 | } |
| 410 | |
| 411 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 412 | _mm256_mask_cmple_epu16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 413 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 2, |
| 414 | __u); |
| 415 | } |
| 416 | |
| 417 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 418 | _mm_cmplt_epi8_mask(__m128i __a, __m128i __b) { |
| 419 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 1, |
| 420 | (__mmask16)-1); |
| 421 | } |
| 422 | |
| 423 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 424 | _mm_mask_cmplt_epi8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 425 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 1, |
| 426 | __u); |
| 427 | } |
| 428 | |
| 429 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 430 | _mm_cmplt_epu8_mask(__m128i __a, __m128i __b) { |
| 431 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 1, |
| 432 | (__mmask16)-1); |
| 433 | } |
| 434 | |
| 435 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 436 | _mm_mask_cmplt_epu8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 437 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 1, |
| 438 | __u); |
| 439 | } |
| 440 | |
| 441 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 442 | _mm256_cmplt_epi8_mask(__m256i __a, __m256i __b) { |
| 443 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 1, |
| 444 | (__mmask32)-1); |
| 445 | } |
| 446 | |
| 447 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 448 | _mm256_mask_cmplt_epi8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 449 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 1, |
| 450 | __u); |
| 451 | } |
| 452 | |
| 453 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 454 | _mm256_cmplt_epu8_mask(__m256i __a, __m256i __b) { |
| 455 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 1, |
| 456 | (__mmask32)-1); |
| 457 | } |
| 458 | |
| 459 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 460 | _mm256_mask_cmplt_epu8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 461 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 1, |
| 462 | __u); |
| 463 | } |
| 464 | |
| 465 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 466 | _mm_cmplt_epi16_mask(__m128i __a, __m128i __b) { |
| 467 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 1, |
| 468 | (__mmask8)-1); |
| 469 | } |
| 470 | |
| 471 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 472 | _mm_mask_cmplt_epi16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 473 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 1, |
| 474 | __u); |
| 475 | } |
| 476 | |
| 477 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 478 | _mm_cmplt_epu16_mask(__m128i __a, __m128i __b) { |
| 479 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 1, |
| 480 | (__mmask8)-1); |
| 481 | } |
| 482 | |
| 483 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 484 | _mm_mask_cmplt_epu16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 485 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 1, |
| 486 | __u); |
| 487 | } |
| 488 | |
| 489 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 490 | _mm256_cmplt_epi16_mask(__m256i __a, __m256i __b) { |
| 491 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 1, |
| 492 | (__mmask16)-1); |
| 493 | } |
| 494 | |
| 495 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 496 | _mm256_mask_cmplt_epi16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 497 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 1, |
| 498 | __u); |
| 499 | } |
| 500 | |
| 501 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 502 | _mm256_cmplt_epu16_mask(__m256i __a, __m256i __b) { |
| 503 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 1, |
| 504 | (__mmask16)-1); |
| 505 | } |
| 506 | |
| 507 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 508 | _mm256_mask_cmplt_epu16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 509 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 1, |
| 510 | __u); |
| 511 | } |
| 512 | |
| 513 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 514 | _mm_cmpneq_epi8_mask(__m128i __a, __m128i __b) { |
| 515 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 4, |
| 516 | (__mmask16)-1); |
| 517 | } |
| 518 | |
| 519 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 520 | _mm_mask_cmpneq_epi8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 521 | return (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)__a, (__v16qi)__b, 4, |
| 522 | __u); |
| 523 | } |
| 524 | |
| 525 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 526 | _mm_cmpneq_epu8_mask(__m128i __a, __m128i __b) { |
| 527 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 4, |
| 528 | (__mmask16)-1); |
| 529 | } |
| 530 | |
| 531 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 532 | _mm_mask_cmpneq_epu8_mask(__mmask16 __u, __m128i __a, __m128i __b) { |
| 533 | return (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)__a, (__v16qi)__b, 4, |
| 534 | __u); |
| 535 | } |
| 536 | |
| 537 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 538 | _mm256_cmpneq_epi8_mask(__m256i __a, __m256i __b) { |
| 539 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 4, |
| 540 | (__mmask32)-1); |
| 541 | } |
| 542 | |
| 543 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 544 | _mm256_mask_cmpneq_epi8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 545 | return (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)__a, (__v32qi)__b, 4, |
| 546 | __u); |
| 547 | } |
| 548 | |
| 549 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 550 | _mm256_cmpneq_epu8_mask(__m256i __a, __m256i __b) { |
| 551 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 4, |
| 552 | (__mmask32)-1); |
| 553 | } |
| 554 | |
| 555 | static __inline__ __mmask32 __attribute__((__always_inline__, __nodebug__)) |
| 556 | _mm256_mask_cmpneq_epu8_mask(__mmask32 __u, __m256i __a, __m256i __b) { |
| 557 | return (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)__a, (__v32qi)__b, 4, |
| 558 | __u); |
| 559 | } |
| 560 | |
| 561 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 562 | _mm_cmpneq_epi16_mask(__m128i __a, __m128i __b) { |
| 563 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 4, |
| 564 | (__mmask8)-1); |
| 565 | } |
| 566 | |
| 567 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 568 | _mm_mask_cmpneq_epi16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 569 | return (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)__a, (__v8hi)__b, 4, |
| 570 | __u); |
| 571 | } |
| 572 | |
| 573 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 574 | _mm_cmpneq_epu16_mask(__m128i __a, __m128i __b) { |
| 575 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 4, |
| 576 | (__mmask8)-1); |
| 577 | } |
| 578 | |
| 579 | static __inline__ __mmask8 __attribute__((__always_inline__, __nodebug__)) |
| 580 | _mm_mask_cmpneq_epu16_mask(__mmask8 __u, __m128i __a, __m128i __b) { |
| 581 | return (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)__a, (__v8hi)__b, 4, |
| 582 | __u); |
| 583 | } |
| 584 | |
| 585 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 586 | _mm256_cmpneq_epi16_mask(__m256i __a, __m256i __b) { |
| 587 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 4, |
| 588 | (__mmask16)-1); |
| 589 | } |
| 590 | |
| 591 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 592 | _mm256_mask_cmpneq_epi16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 593 | return (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)__a, (__v16hi)__b, 4, |
| 594 | __u); |
| 595 | } |
| 596 | |
| 597 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 598 | _mm256_cmpneq_epu16_mask(__m256i __a, __m256i __b) { |
| 599 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 4, |
| 600 | (__mmask16)-1); |
| 601 | } |
| 602 | |
| 603 | static __inline__ __mmask16 __attribute__((__always_inline__, __nodebug__)) |
| 604 | _mm256_mask_cmpneq_epu16_mask(__mmask16 __u, __m256i __a, __m256i __b) { |
| 605 | return (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)__a, (__v16hi)__b, 4, |
| 606 | __u); |
| 607 | } |
| 608 | |
| 609 | #define _mm_cmp_epi8_mask(a, b, p) __extension__ ({ \ |
| 610 | (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \ |
| 611 | (__v16qi)(__m128i)(b), \ |
| 612 | (p), (__mmask16)-1); }) |
| 613 | |
| 614 | #define _mm_mask_cmp_epi8_mask(m, a, b, p) __extension__ ({ \ |
| 615 | (__mmask16)__builtin_ia32_cmpb128_mask((__v16qi)(__m128i)(a), \ |
| 616 | (__v16qi)(__m128i)(b), \ |
| 617 | (p), (__mmask16)(m)); }) |
| 618 | |
| 619 | #define _mm_cmp_epu8_mask(a, b, p) __extension__ ({ \ |
| 620 | (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \ |
| 621 | (__v16qi)(__m128i)(b), \ |
| 622 | (p), (__mmask16)-1); }) |
| 623 | |
| 624 | #define _mm_mask_cmp_epu8_mask(m, a, b, p) __extension__ ({ \ |
| 625 | (__mmask16)__builtin_ia32_ucmpb128_mask((__v16qi)(__m128i)(a), \ |
| 626 | (__v16qi)(__m128i)(b), \ |
| 627 | (p), (__mmask16)(m)); }) |
| 628 | |
| 629 | #define _mm256_cmp_epi8_mask(a, b, p) __extension__ ({ \ |
| 630 | (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \ |
| 631 | (__v32qi)(__m256i)(b), \ |
| 632 | (p), (__mmask32)-1); }) |
| 633 | |
| 634 | #define _mm256_mask_cmp_epi8_mask(m, a, b, p) __extension__ ({ \ |
| 635 | (__mmask32)__builtin_ia32_cmpb256_mask((__v32qi)(__m256i)(a), \ |
| 636 | (__v32qi)(__m256i)(b), \ |
| 637 | (p), (__mmask32)(m)); }) |
| 638 | |
| 639 | #define _mm256_cmp_epu8_mask(a, b, p) __extension__ ({ \ |
| 640 | (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \ |
| 641 | (__v32qi)(__m256i)(b), \ |
| 642 | (p), (__mmask32)-1); }) |
| 643 | |
| 644 | #define _mm256_mask_cmp_epu8_mask(m, a, b, p) __extension__ ({ \ |
| 645 | (__mmask32)__builtin_ia32_ucmpb256_mask((__v32qi)(__m256i)(a), \ |
| 646 | (__v32qi)(__m256i)(b), \ |
| 647 | (p), (__mmask32)(m)); }) |
| 648 | |
| 649 | #define _mm_cmp_epi16_mask(a, b, p) __extension__ ({ \ |
| 650 | (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \ |
| 651 | (__v8hi)(__m128i)(b), \ |
| 652 | (p), (__mmask8)-1); }) |
| 653 | |
| 654 | #define _mm_mask_cmp_epi16_mask(m, a, b, p) __extension__ ({ \ |
| 655 | (__mmask8)__builtin_ia32_cmpw128_mask((__v8hi)(__m128i)(a), \ |
| 656 | (__v8hi)(__m128i)(b), \ |
| 657 | (p), (__mmask8)(m)); }) |
| 658 | |
| 659 | #define _mm_cmp_epu16_mask(a, b, p) __extension__ ({ \ |
| 660 | (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \ |
| 661 | (__v8hi)(__m128i)(b), \ |
| 662 | (p), (__mmask8)-1); }) |
| 663 | |
| 664 | #define _mm_mask_cmp_epu16_mask(m, a, b, p) __extension__ ({ \ |
| 665 | (__mmask8)__builtin_ia32_ucmpw128_mask((__v8hi)(__m128i)(a), \ |
| 666 | (__v8hi)(__m128i)(b), \ |
| 667 | (p), (__mmask8)(m)); }) |
| 668 | |
| 669 | #define _mm256_cmp_epi16_mask(a, b, p) __extension__ ({ \ |
| 670 | (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \ |
| 671 | (__v16hi)(__m256i)(b), \ |
| 672 | (p), (__mmask16)-1); }) |
| 673 | |
| 674 | #define _mm256_mask_cmp_epi16_mask(m, a, b, p) __extension__ ({ \ |
| 675 | (__mmask16)__builtin_ia32_cmpw256_mask((__v16hi)(__m256i)(a), \ |
| 676 | (__v16hi)(__m256i)(b), \ |
| 677 | (p), (__mmask16)(m)); }) |
| 678 | |
| 679 | #define _mm256_cmp_epu16_mask(a, b, p) __extension__ ({ \ |
| 680 | (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \ |
| 681 | (__v16hi)(__m256i)(b), \ |
| 682 | (p), (__mmask16)-1); }) |
| 683 | |
| 684 | #define _mm256_mask_cmp_epu16_mask(m, a, b, p) __extension__ ({ \ |
| 685 | (__mmask16)__builtin_ia32_ucmpw256_mask((__v16hi)(__m256i)(a), \ |
| 686 | (__v16hi)(__m256i)(b), \ |
| 687 | (p), (__mmask16)(m)); }) |
| 688 | |
Stephen Hines | ee4ca28 | 2014-12-02 17:05:12 -0800 | [diff] [blame] | 689 | #endif /* __AVX512VLBWINTRIN_H */ |