| 1 | /*===------ avx10_2_512satcvtintrin.h - AVX10_2_512SATCVT intrinsics -------=== |
| 2 | * |
| 3 | * Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions. |
| 4 | * See https://llvm.org/LICENSE.txt for license information. |
| 5 | * SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception |
| 6 | * |
| 7 | *===-----------------------------------------------------------------------=== |
| 8 | */ |
| 9 | #ifndef __IMMINTRIN_H |
| 10 | #error \ |
| 11 | "Never use <avx10_2_512satcvtintrin.h> directly; include <immintrin.h> instead." |
| 12 | #endif // __IMMINTRIN_H |
| 13 | |
| 14 | #ifndef __AVX10_2_512SATCVTINTRIN_H |
| 15 | #define __AVX10_2_512SATCVTINTRIN_H |
| 16 | |
| 17 | #define _mm512_ipcvts_bf16_epi8(A) \ |
| 18 | ((__m512i)__builtin_ia32_vcvtbf162ibs512((__v32bf)(__m512bh)(A))) |
| 19 | |
| 20 | #define _mm512_mask_ipcvts_bf16_epi8(W, U, A) \ |
| 21 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 22 | (__v32hi)_mm512_ipcvts_bf16_epi8(A), \ |
| 23 | (__v32hi)(__m512i)(W))) |
| 24 | |
| 25 | #define _mm512_maskz_ipcvts_bf16_epi8(U, A) \ |
| 26 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 27 | (__v32hi)_mm512_ipcvts_bf16_epi8(A), \ |
| 28 | (__v32hi)_mm512_setzero_si512())) |
| 29 | |
| 30 | #define _mm512_ipcvts_bf16_epu8(A) \ |
| 31 | ((__m512i)__builtin_ia32_vcvtbf162iubs512((__v32bf)(__m512bh)(A))) |
| 32 | |
| 33 | #define _mm512_mask_ipcvts_bf16_epu8(W, U, A) \ |
| 34 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 35 | (__v32hi)_mm512_ipcvts_bf16_epu8(A), \ |
| 36 | (__v32hi)(__m512i)(W))) |
| 37 | |
| 38 | #define _mm512_maskz_ipcvts_bf16_epu8(U, A) \ |
| 39 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 40 | (__v32hi)_mm512_ipcvts_bf16_epu8(A), \ |
| 41 | (__v32hi)_mm512_setzero_si512())) |
| 42 | |
| 43 | #define _mm512_ipcvtts_bf16_epi8(A) \ |
| 44 | ((__m512i)__builtin_ia32_vcvttbf162ibs512((__v32bf)(__m512bh)(A))) |
| 45 | |
| 46 | #define _mm512_mask_ipcvtts_bf16_epi8(W, U, A) \ |
| 47 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 48 | (__v32hi)_mm512_ipcvtts_bf16_epi8(A), \ |
| 49 | (__v32hi)(__m512i)(W))) |
| 50 | |
| 51 | #define _mm512_maskz_ipcvtts_bf16_epi8(U, A) \ |
| 52 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 53 | (__v32hi)_mm512_ipcvtts_bf16_epi8(A), \ |
| 54 | (__v32hi)_mm512_setzero_si512())) |
| 55 | |
| 56 | #define _mm512_ipcvtts_bf16_epu8(A) \ |
| 57 | ((__m512i)__builtin_ia32_vcvttbf162iubs512((__v32bf)(__m512bh)(A))) |
| 58 | |
| 59 | #define _mm512_mask_ipcvtts_bf16_epu8(W, U, A) \ |
| 60 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 61 | (__v32hi)_mm512_ipcvtts_bf16_epu8(A), \ |
| 62 | (__v32hi)(__m512i)(W))) |
| 63 | |
| 64 | #define _mm512_maskz_ipcvtts_bf16_epu8(U, A) \ |
| 65 | ((__m512i)__builtin_ia32_selectw_512((__mmask32)(U), \ |
| 66 | (__v32hi)_mm512_ipcvtts_bf16_epu8(A), \ |
| 67 | (__v32hi)_mm512_setzero_si512())) |
| 68 | |
| 69 | #define _mm512_ipcvts_ph_epi8(A) \ |
| 70 | ((__m512i)__builtin_ia32_vcvtph2ibs512_mask( \ |
| 71 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 72 | _MM_FROUND_CUR_DIRECTION)) |
| 73 | |
| 74 | #define _mm512_mask_ipcvts_ph_epi8(W, U, A) \ |
| 75 | ((__m512i)__builtin_ia32_vcvtph2ibs512_mask((__v32hf)(__m512h)(A), \ |
| 76 | (__v32hu)(W), (__mmask32)(U), \ |
| 77 | _MM_FROUND_CUR_DIRECTION)) |
| 78 | |
| 79 | #define _mm512_maskz_ipcvts_ph_epi8(U, A) \ |
| 80 | ((__m512i)__builtin_ia32_vcvtph2ibs512_mask( \ |
| 81 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 82 | _MM_FROUND_CUR_DIRECTION)) |
| 83 | |
| 84 | #define _mm512_ipcvts_roundph_epi8(A, R) \ |
| 85 | ((__m512i)__builtin_ia32_vcvtph2ibs512_mask((__v32hf)(__m512h)(A), \ |
| 86 | (__v32hu)_mm512_setzero_si512(), \ |
| 87 | (__mmask32) - 1, (const int)R)) |
| 88 | |
| 89 | #define _mm512_mask_ipcvts_roundph_epi8(W, U, A, R) \ |
| 90 | ((__m512i)__builtin_ia32_vcvtph2ibs512_mask( \ |
| 91 | (__v32hf)(__m512h)(A), (__v32hu)(W), (__mmask32)(U), (const int)R)) |
| 92 | |
| 93 | #define _mm512_maskz_ipcvts_roundph_epi8(U, A, R) \ |
| 94 | ((__m512i)__builtin_ia32_vcvtph2ibs512_mask((__v32hf)(__m512h)(A), \ |
| 95 | (__v32hu)_mm512_setzero_si512(), \ |
| 96 | (__mmask32)(U), (const int)R)) |
| 97 | |
| 98 | #define _mm512_ipcvts_ph_epu8(A) \ |
| 99 | ((__m512i)__builtin_ia32_vcvtph2iubs512_mask( \ |
| 100 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 101 | _MM_FROUND_CUR_DIRECTION)) |
| 102 | |
| 103 | #define _mm512_mask_ipcvts_ph_epu8(W, U, A) \ |
| 104 | ((__m512i)__builtin_ia32_vcvtph2iubs512_mask((__v32hf)(__m512h)(A), \ |
| 105 | (__v32hu)(W), (__mmask32)(U), \ |
| 106 | _MM_FROUND_CUR_DIRECTION)) |
| 107 | |
| 108 | #define _mm512_maskz_ipcvts_ph_epu8(U, A) \ |
| 109 | ((__m512i)__builtin_ia32_vcvtph2iubs512_mask( \ |
| 110 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 111 | _MM_FROUND_CUR_DIRECTION)) |
| 112 | |
| 113 | #define _mm512_ipcvts_roundph_epu8(A, R) \ |
| 114 | ((__m512i)__builtin_ia32_vcvtph2iubs512_mask( \ |
| 115 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 116 | (const int)R)) |
| 117 | |
| 118 | #define _mm512_mask_ipcvts_roundph_epu8(W, U, A, R) \ |
| 119 | ((__m512i)__builtin_ia32_vcvtph2iubs512_mask( \ |
| 120 | (__v32hf)(__m512h)(A), (__v32hu)(W), (__mmask32)(U), (const int)R)) |
| 121 | |
| 122 | #define _mm512_maskz_ipcvts_roundph_epu8(U, A, R) \ |
| 123 | ((__m512i)__builtin_ia32_vcvtph2iubs512_mask( \ |
| 124 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 125 | (const int)R)) |
| 126 | |
| 127 | #define _mm512_ipcvts_ps_epi8(A) \ |
| 128 | ((__m512i)__builtin_ia32_vcvtps2ibs512_mask( \ |
| 129 | (__v16sf)(__m512)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 130 | _MM_FROUND_CUR_DIRECTION)) |
| 131 | |
| 132 | #define _mm512_mask_ipcvts_ps_epi8(W, U, A) \ |
| 133 | ((__m512i)__builtin_ia32_vcvtps2ibs512_mask((__v16sf)(__m512)(A), \ |
| 134 | (__v16su)(W), (__mmask16)(U), \ |
| 135 | _MM_FROUND_CUR_DIRECTION)) |
| 136 | |
| 137 | #define _mm512_maskz_ipcvts_ps_epi8(U, A) \ |
| 138 | ((__m512i)__builtin_ia32_vcvtps2ibs512_mask( \ |
| 139 | (__v16sf)(__m512)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 140 | _MM_FROUND_CUR_DIRECTION)) |
| 141 | |
| 142 | #define _mm512_ipcvts_roundps_epi8(A, R) \ |
| 143 | ((__m512i)__builtin_ia32_vcvtps2ibs512_mask((__v16sf)(__m512)(A), \ |
| 144 | (__v16su)_mm512_setzero_si512(), \ |
| 145 | (__mmask16) - 1, (const int)R)) |
| 146 | |
| 147 | #define _mm512_mask_ipcvts_roundps_epi8(W, U, A, R) \ |
| 148 | ((__m512i)__builtin_ia32_vcvtps2ibs512_mask( \ |
| 149 | (__v16sf)(__m512)(A), (__v16su)(W), (__mmask16)(U), (const int)R)) |
| 150 | |
| 151 | #define _mm512_maskz_ipcvts_roundps_epi8(U, A, R) \ |
| 152 | ((__m512i)__builtin_ia32_vcvtps2ibs512_mask((__v16sf)(__m512)(A), \ |
| 153 | (__v16su)_mm512_setzero_si512(), \ |
| 154 | (__mmask16)(U), (const int)R)) |
| 155 | |
| 156 | #define _mm512_ipcvts_ps_epu8(A) \ |
| 157 | ((__m512i)__builtin_ia32_vcvtps2iubs512_mask( \ |
| 158 | (__v16sf)(__m512)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 159 | _MM_FROUND_CUR_DIRECTION)) |
| 160 | |
| 161 | #define _mm512_mask_ipcvts_ps_epu8(W, U, A) \ |
| 162 | ((__m512i)__builtin_ia32_vcvtps2iubs512_mask((__v16sf)(__m512)(A), \ |
| 163 | (__v16su)(W), (__mmask16)(U), \ |
| 164 | _MM_FROUND_CUR_DIRECTION)) |
| 165 | |
| 166 | #define _mm512_maskz_ipcvts_ps_epu8(U, A) \ |
| 167 | ((__m512i)__builtin_ia32_vcvtps2iubs512_mask( \ |
| 168 | (__v16sf)(__m512)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 169 | _MM_FROUND_CUR_DIRECTION)) |
| 170 | |
| 171 | #define _mm512_ipcvts_roundps_epu8(A, R) \ |
| 172 | ((__m512i)__builtin_ia32_vcvtps2iubs512_mask( \ |
| 173 | (__v16sf)(__m512)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 174 | (const int)R)) |
| 175 | |
| 176 | #define _mm512_mask_ipcvts_roundps_epu8(W, U, A, R) \ |
| 177 | ((__m512i)__builtin_ia32_vcvtps2iubs512_mask( \ |
| 178 | (__v16sf)(__m512)(A), (__v16su)(W), (__mmask16)(U), (const int)R)) |
| 179 | |
| 180 | #define _mm512_maskz_ipcvts_roundps_epu8(U, A, R) \ |
| 181 | ((__m512i)__builtin_ia32_vcvtps2iubs512_mask( \ |
| 182 | (__v16sf)(__m512)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 183 | (const int)R)) |
| 184 | |
| 185 | #define _mm512_ipcvtts_ph_epi8(A) \ |
| 186 | ((__m512i)__builtin_ia32_vcvttph2ibs512_mask( \ |
| 187 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 188 | _MM_FROUND_CUR_DIRECTION)) |
| 189 | |
| 190 | #define _mm512_mask_ipcvtts_ph_epi8(W, U, A) \ |
| 191 | ((__m512i)__builtin_ia32_vcvttph2ibs512_mask((__v32hf)(__m512h)(A), \ |
| 192 | (__v32hu)(W), (__mmask32)(U), \ |
| 193 | _MM_FROUND_CUR_DIRECTION)) |
| 194 | |
| 195 | #define _mm512_maskz_ipcvtts_ph_epi8(U, A) \ |
| 196 | ((__m512i)__builtin_ia32_vcvttph2ibs512_mask( \ |
| 197 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 198 | _MM_FROUND_CUR_DIRECTION)) |
| 199 | |
| 200 | #define _mm512_ipcvtts_roundph_epi8(A, S) \ |
| 201 | ((__m512i)__builtin_ia32_vcvttph2ibs512_mask( \ |
| 202 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 203 | S)) |
| 204 | |
| 205 | #define _mm512_mask_ipcvtts_roundph_epi8(W, U, A, S) \ |
| 206 | ((__m512i)__builtin_ia32_vcvttph2ibs512_mask( \ |
| 207 | (__v32hf)(__m512h)(A), (__v32hu)(W), (__mmask32)(U), S)) |
| 208 | |
| 209 | #define _mm512_maskz_ipcvtts_roundph_epi8(U, A, S) \ |
| 210 | ((__m512i)__builtin_ia32_vcvttph2ibs512_mask( \ |
| 211 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 212 | S)) |
| 213 | |
| 214 | #define _mm512_ipcvtts_ph_epu8(A) \ |
| 215 | ((__m512i)__builtin_ia32_vcvttph2iubs512_mask( \ |
| 216 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 217 | _MM_FROUND_CUR_DIRECTION)) |
| 218 | |
| 219 | #define _mm512_mask_ipcvtts_ph_epu8(W, U, A) \ |
| 220 | ((__m512i)__builtin_ia32_vcvttph2iubs512_mask((__v32hf)(__m512h)(A), \ |
| 221 | (__v32hu)(W), (__mmask32)(U), \ |
| 222 | _MM_FROUND_CUR_DIRECTION)) |
| 223 | |
| 224 | #define _mm512_maskz_ipcvtts_ph_epu8(U, A) \ |
| 225 | ((__m512i)__builtin_ia32_vcvttph2iubs512_mask( \ |
| 226 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 227 | _MM_FROUND_CUR_DIRECTION)) |
| 228 | |
| 229 | #define _mm512_ipcvtts_roundph_epu8(A, S) \ |
| 230 | ((__m512i)__builtin_ia32_vcvttph2iubs512_mask( \ |
| 231 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32) - 1, \ |
| 232 | S)) |
| 233 | |
| 234 | #define _mm512_mask_ipcvtts_roundph_epu8(W, U, A, S) \ |
| 235 | ((__m512i)__builtin_ia32_vcvttph2iubs512_mask( \ |
| 236 | (__v32hf)(__m512h)(A), (__v32hu)(W), (__mmask32)(U), S)) |
| 237 | |
| 238 | #define _mm512_maskz_ipcvtts_roundph_epu8(U, A, S) \ |
| 239 | ((__m512i)__builtin_ia32_vcvttph2iubs512_mask( \ |
| 240 | (__v32hf)(__m512h)(A), (__v32hu)_mm512_setzero_si512(), (__mmask32)(U), \ |
| 241 | S)) |
| 242 | |
| 243 | #define _mm512_ipcvtts_ps_epi8(A) \ |
| 244 | ((__m512i)__builtin_ia32_vcvttps2ibs512_mask( \ |
| 245 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 246 | _MM_FROUND_CUR_DIRECTION)) |
| 247 | |
| 248 | #define _mm512_mask_ipcvtts_ps_epi8(W, U, A) \ |
| 249 | ((__m512i)__builtin_ia32_vcvttps2ibs512_mask((__v16sf)(__m512h)(A), \ |
| 250 | (__v16su)(W), (__mmask16)(U), \ |
| 251 | _MM_FROUND_CUR_DIRECTION)) |
| 252 | |
| 253 | #define _mm512_maskz_ipcvtts_ps_epi8(U, A) \ |
| 254 | ((__m512i)__builtin_ia32_vcvttps2ibs512_mask( \ |
| 255 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 256 | _MM_FROUND_CUR_DIRECTION)) |
| 257 | |
| 258 | #define _mm512_ipcvtts_roundps_epi8(A, S) \ |
| 259 | ((__m512i)__builtin_ia32_vcvttps2ibs512_mask( \ |
| 260 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 261 | S)) |
| 262 | |
| 263 | #define _mm512_mask_ipcvtts_roundps_epi8(W, U, A, S) \ |
| 264 | ((__m512i)__builtin_ia32_vcvttps2ibs512_mask( \ |
| 265 | (__v16sf)(__m512h)(A), (__v16su)(W), (__mmask16)(U), S)) |
| 266 | |
| 267 | #define _mm512_maskz_ipcvtts_roundps_epi8(U, A, S) \ |
| 268 | ((__m512i)__builtin_ia32_vcvttps2ibs512_mask( \ |
| 269 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 270 | S)) |
| 271 | |
| 272 | #define _mm512_ipcvtts_ps_epu8(A) \ |
| 273 | ((__m512i)__builtin_ia32_vcvttps2iubs512_mask( \ |
| 274 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 275 | _MM_FROUND_CUR_DIRECTION)) |
| 276 | |
| 277 | #define _mm512_mask_ipcvtts_ps_epu8(W, U, A) \ |
| 278 | ((__m512i)__builtin_ia32_vcvttps2iubs512_mask((__v16sf)(__m512h)(A), \ |
| 279 | (__v16su)(W), (__mmask16)(U), \ |
| 280 | _MM_FROUND_CUR_DIRECTION)) |
| 281 | |
| 282 | #define _mm512_maskz_ipcvtts_ps_epu8(U, A) \ |
| 283 | ((__m512i)__builtin_ia32_vcvttps2iubs512_mask( \ |
| 284 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 285 | _MM_FROUND_CUR_DIRECTION)) |
| 286 | |
| 287 | #define _mm512_ipcvtts_roundps_epu8(A, S) \ |
| 288 | ((__m512i)__builtin_ia32_vcvttps2iubs512_mask( \ |
| 289 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16) - 1, \ |
| 290 | S)) |
| 291 | |
| 292 | #define _mm512_mask_ipcvtts_roundps_epu8(W, U, A, S) \ |
| 293 | ((__m512i)__builtin_ia32_vcvttps2iubs512_mask( \ |
| 294 | (__v16sf)(__m512h)(A), (__v16su)(W), (__mmask16)(U), S)) |
| 295 | |
| 296 | #define _mm512_maskz_ipcvtts_roundps_epu8(U, A, S) \ |
| 297 | ((__m512i)__builtin_ia32_vcvttps2iubs512_mask( \ |
| 298 | (__v16sf)(__m512h)(A), (__v16su)_mm512_setzero_si512(), (__mmask16)(U), \ |
| 299 | S)) |
| 300 | |
| 301 | #endif // __AVX10_2_512SATCVTINTRIN_H |