Details | Last modification | View Log | RSS feed
| Rev | Author | Line No. | Line |
|---|---|---|---|
| 14 | pmbaty | 1 | /*===---- xopintrin.h - XOP 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 __X86INTRIN_H |
||
| 11 | #error "Never use <xopintrin.h> directly; include <x86intrin.h> instead." |
||
| 12 | #endif |
||
| 13 | |||
| 14 | #ifndef __XOPINTRIN_H |
||
| 15 | #define __XOPINTRIN_H |
||
| 16 | |||
| 17 | #include <fma4intrin.h> |
||
| 18 | |||
| 19 | /* Define the default attributes for the functions in this file. */ |
||
| 20 | #define __DEFAULT_FN_ATTRS __attribute__((__always_inline__, __nodebug__, __target__("xop"), __min_vector_width__(128))) |
||
| 21 | #define __DEFAULT_FN_ATTRS256 __attribute__((__always_inline__, __nodebug__, __target__("xop"), __min_vector_width__(256))) |
||
| 22 | |||
| 23 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 24 | _mm_maccs_epi16(__m128i __A, __m128i __B, __m128i __C) |
||
| 25 | { |
||
| 26 | return (__m128i)__builtin_ia32_vpmacssww((__v8hi)__A, (__v8hi)__B, (__v8hi)__C); |
||
| 27 | } |
||
| 28 | |||
| 29 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 30 | _mm_macc_epi16(__m128i __A, __m128i __B, __m128i __C) |
||
| 31 | { |
||
| 32 | return (__m128i)__builtin_ia32_vpmacsww((__v8hi)__A, (__v8hi)__B, (__v8hi)__C); |
||
| 33 | } |
||
| 34 | |||
| 35 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 36 | _mm_maccsd_epi16(__m128i __A, __m128i __B, __m128i __C) |
||
| 37 | { |
||
| 38 | return (__m128i)__builtin_ia32_vpmacsswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C); |
||
| 39 | } |
||
| 40 | |||
| 41 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 42 | _mm_maccd_epi16(__m128i __A, __m128i __B, __m128i __C) |
||
| 43 | { |
||
| 44 | return (__m128i)__builtin_ia32_vpmacswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C); |
||
| 45 | } |
||
| 46 | |||
| 47 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 48 | _mm_maccs_epi32(__m128i __A, __m128i __B, __m128i __C) |
||
| 49 | { |
||
| 50 | return (__m128i)__builtin_ia32_vpmacssdd((__v4si)__A, (__v4si)__B, (__v4si)__C); |
||
| 51 | } |
||
| 52 | |||
| 53 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 54 | _mm_macc_epi32(__m128i __A, __m128i __B, __m128i __C) |
||
| 55 | { |
||
| 56 | return (__m128i)__builtin_ia32_vpmacsdd((__v4si)__A, (__v4si)__B, (__v4si)__C); |
||
| 57 | } |
||
| 58 | |||
| 59 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 60 | _mm_maccslo_epi32(__m128i __A, __m128i __B, __m128i __C) |
||
| 61 | { |
||
| 62 | return (__m128i)__builtin_ia32_vpmacssdql((__v4si)__A, (__v4si)__B, (__v2di)__C); |
||
| 63 | } |
||
| 64 | |||
| 65 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 66 | _mm_macclo_epi32(__m128i __A, __m128i __B, __m128i __C) |
||
| 67 | { |
||
| 68 | return (__m128i)__builtin_ia32_vpmacsdql((__v4si)__A, (__v4si)__B, (__v2di)__C); |
||
| 69 | } |
||
| 70 | |||
| 71 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 72 | _mm_maccshi_epi32(__m128i __A, __m128i __B, __m128i __C) |
||
| 73 | { |
||
| 74 | return (__m128i)__builtin_ia32_vpmacssdqh((__v4si)__A, (__v4si)__B, (__v2di)__C); |
||
| 75 | } |
||
| 76 | |||
| 77 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 78 | _mm_macchi_epi32(__m128i __A, __m128i __B, __m128i __C) |
||
| 79 | { |
||
| 80 | return (__m128i)__builtin_ia32_vpmacsdqh((__v4si)__A, (__v4si)__B, (__v2di)__C); |
||
| 81 | } |
||
| 82 | |||
| 83 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 84 | _mm_maddsd_epi16(__m128i __A, __m128i __B, __m128i __C) |
||
| 85 | { |
||
| 86 | return (__m128i)__builtin_ia32_vpmadcsswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C); |
||
| 87 | } |
||
| 88 | |||
| 89 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 90 | _mm_maddd_epi16(__m128i __A, __m128i __B, __m128i __C) |
||
| 91 | { |
||
| 92 | return (__m128i)__builtin_ia32_vpmadcswd((__v8hi)__A, (__v8hi)__B, (__v4si)__C); |
||
| 93 | } |
||
| 94 | |||
| 95 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 96 | _mm_haddw_epi8(__m128i __A) |
||
| 97 | { |
||
| 98 | return (__m128i)__builtin_ia32_vphaddbw((__v16qi)__A); |
||
| 99 | } |
||
| 100 | |||
| 101 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 102 | _mm_haddd_epi8(__m128i __A) |
||
| 103 | { |
||
| 104 | return (__m128i)__builtin_ia32_vphaddbd((__v16qi)__A); |
||
| 105 | } |
||
| 106 | |||
| 107 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 108 | _mm_haddq_epi8(__m128i __A) |
||
| 109 | { |
||
| 110 | return (__m128i)__builtin_ia32_vphaddbq((__v16qi)__A); |
||
| 111 | } |
||
| 112 | |||
| 113 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 114 | _mm_haddd_epi16(__m128i __A) |
||
| 115 | { |
||
| 116 | return (__m128i)__builtin_ia32_vphaddwd((__v8hi)__A); |
||
| 117 | } |
||
| 118 | |||
| 119 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 120 | _mm_haddq_epi16(__m128i __A) |
||
| 121 | { |
||
| 122 | return (__m128i)__builtin_ia32_vphaddwq((__v8hi)__A); |
||
| 123 | } |
||
| 124 | |||
| 125 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 126 | _mm_haddq_epi32(__m128i __A) |
||
| 127 | { |
||
| 128 | return (__m128i)__builtin_ia32_vphadddq((__v4si)__A); |
||
| 129 | } |
||
| 130 | |||
| 131 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 132 | _mm_haddw_epu8(__m128i __A) |
||
| 133 | { |
||
| 134 | return (__m128i)__builtin_ia32_vphaddubw((__v16qi)__A); |
||
| 135 | } |
||
| 136 | |||
| 137 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 138 | _mm_haddd_epu8(__m128i __A) |
||
| 139 | { |
||
| 140 | return (__m128i)__builtin_ia32_vphaddubd((__v16qi)__A); |
||
| 141 | } |
||
| 142 | |||
| 143 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 144 | _mm_haddq_epu8(__m128i __A) |
||
| 145 | { |
||
| 146 | return (__m128i)__builtin_ia32_vphaddubq((__v16qi)__A); |
||
| 147 | } |
||
| 148 | |||
| 149 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 150 | _mm_haddd_epu16(__m128i __A) |
||
| 151 | { |
||
| 152 | return (__m128i)__builtin_ia32_vphadduwd((__v8hi)__A); |
||
| 153 | } |
||
| 154 | |||
| 155 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 156 | _mm_haddq_epu16(__m128i __A) |
||
| 157 | { |
||
| 158 | return (__m128i)__builtin_ia32_vphadduwq((__v8hi)__A); |
||
| 159 | } |
||
| 160 | |||
| 161 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 162 | _mm_haddq_epu32(__m128i __A) |
||
| 163 | { |
||
| 164 | return (__m128i)__builtin_ia32_vphaddudq((__v4si)__A); |
||
| 165 | } |
||
| 166 | |||
| 167 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 168 | _mm_hsubw_epi8(__m128i __A) |
||
| 169 | { |
||
| 170 | return (__m128i)__builtin_ia32_vphsubbw((__v16qi)__A); |
||
| 171 | } |
||
| 172 | |||
| 173 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 174 | _mm_hsubd_epi16(__m128i __A) |
||
| 175 | { |
||
| 176 | return (__m128i)__builtin_ia32_vphsubwd((__v8hi)__A); |
||
| 177 | } |
||
| 178 | |||
| 179 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 180 | _mm_hsubq_epi32(__m128i __A) |
||
| 181 | { |
||
| 182 | return (__m128i)__builtin_ia32_vphsubdq((__v4si)__A); |
||
| 183 | } |
||
| 184 | |||
| 185 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 186 | _mm_cmov_si128(__m128i __A, __m128i __B, __m128i __C) |
||
| 187 | { |
||
| 188 | return (__m128i)(((__v2du)__A & (__v2du)__C) | ((__v2du)__B & ~(__v2du)__C)); |
||
| 189 | } |
||
| 190 | |||
| 191 | static __inline__ __m256i __DEFAULT_FN_ATTRS256 |
||
| 192 | _mm256_cmov_si256(__m256i __A, __m256i __B, __m256i __C) |
||
| 193 | { |
||
| 194 | return (__m256i)(((__v4du)__A & (__v4du)__C) | ((__v4du)__B & ~(__v4du)__C)); |
||
| 195 | } |
||
| 196 | |||
| 197 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 198 | _mm_perm_epi8(__m128i __A, __m128i __B, __m128i __C) |
||
| 199 | { |
||
| 200 | return (__m128i)__builtin_ia32_vpperm((__v16qi)__A, (__v16qi)__B, (__v16qi)__C); |
||
| 201 | } |
||
| 202 | |||
| 203 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 204 | _mm_rot_epi8(__m128i __A, __m128i __B) |
||
| 205 | { |
||
| 206 | return (__m128i)__builtin_ia32_vprotb((__v16qi)__A, (__v16qi)__B); |
||
| 207 | } |
||
| 208 | |||
| 209 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 210 | _mm_rot_epi16(__m128i __A, __m128i __B) |
||
| 211 | { |
||
| 212 | return (__m128i)__builtin_ia32_vprotw((__v8hi)__A, (__v8hi)__B); |
||
| 213 | } |
||
| 214 | |||
| 215 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 216 | _mm_rot_epi32(__m128i __A, __m128i __B) |
||
| 217 | { |
||
| 218 | return (__m128i)__builtin_ia32_vprotd((__v4si)__A, (__v4si)__B); |
||
| 219 | } |
||
| 220 | |||
| 221 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 222 | _mm_rot_epi64(__m128i __A, __m128i __B) |
||
| 223 | { |
||
| 224 | return (__m128i)__builtin_ia32_vprotq((__v2di)__A, (__v2di)__B); |
||
| 225 | } |
||
| 226 | |||
| 227 | #define _mm_roti_epi8(A, N) \ |
||
| 228 | ((__m128i)__builtin_ia32_vprotbi((__v16qi)(__m128i)(A), (N))) |
||
| 229 | |||
| 230 | #define _mm_roti_epi16(A, N) \ |
||
| 231 | ((__m128i)__builtin_ia32_vprotwi((__v8hi)(__m128i)(A), (N))) |
||
| 232 | |||
| 233 | #define _mm_roti_epi32(A, N) \ |
||
| 234 | ((__m128i)__builtin_ia32_vprotdi((__v4si)(__m128i)(A), (N))) |
||
| 235 | |||
| 236 | #define _mm_roti_epi64(A, N) \ |
||
| 237 | ((__m128i)__builtin_ia32_vprotqi((__v2di)(__m128i)(A), (N))) |
||
| 238 | |||
| 239 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 240 | _mm_shl_epi8(__m128i __A, __m128i __B) |
||
| 241 | { |
||
| 242 | return (__m128i)__builtin_ia32_vpshlb((__v16qi)__A, (__v16qi)__B); |
||
| 243 | } |
||
| 244 | |||
| 245 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 246 | _mm_shl_epi16(__m128i __A, __m128i __B) |
||
| 247 | { |
||
| 248 | return (__m128i)__builtin_ia32_vpshlw((__v8hi)__A, (__v8hi)__B); |
||
| 249 | } |
||
| 250 | |||
| 251 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 252 | _mm_shl_epi32(__m128i __A, __m128i __B) |
||
| 253 | { |
||
| 254 | return (__m128i)__builtin_ia32_vpshld((__v4si)__A, (__v4si)__B); |
||
| 255 | } |
||
| 256 | |||
| 257 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 258 | _mm_shl_epi64(__m128i __A, __m128i __B) |
||
| 259 | { |
||
| 260 | return (__m128i)__builtin_ia32_vpshlq((__v2di)__A, (__v2di)__B); |
||
| 261 | } |
||
| 262 | |||
| 263 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 264 | _mm_sha_epi8(__m128i __A, __m128i __B) |
||
| 265 | { |
||
| 266 | return (__m128i)__builtin_ia32_vpshab((__v16qi)__A, (__v16qi)__B); |
||
| 267 | } |
||
| 268 | |||
| 269 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 270 | _mm_sha_epi16(__m128i __A, __m128i __B) |
||
| 271 | { |
||
| 272 | return (__m128i)__builtin_ia32_vpshaw((__v8hi)__A, (__v8hi)__B); |
||
| 273 | } |
||
| 274 | |||
| 275 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 276 | _mm_sha_epi32(__m128i __A, __m128i __B) |
||
| 277 | { |
||
| 278 | return (__m128i)__builtin_ia32_vpshad((__v4si)__A, (__v4si)__B); |
||
| 279 | } |
||
| 280 | |||
| 281 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 282 | _mm_sha_epi64(__m128i __A, __m128i __B) |
||
| 283 | { |
||
| 284 | return (__m128i)__builtin_ia32_vpshaq((__v2di)__A, (__v2di)__B); |
||
| 285 | } |
||
| 286 | |||
| 287 | #define _mm_com_epu8(A, B, N) \ |
||
| 288 | ((__m128i)__builtin_ia32_vpcomub((__v16qi)(__m128i)(A), \ |
||
| 289 | (__v16qi)(__m128i)(B), (N))) |
||
| 290 | |||
| 291 | #define _mm_com_epu16(A, B, N) \ |
||
| 292 | ((__m128i)__builtin_ia32_vpcomuw((__v8hi)(__m128i)(A), \ |
||
| 293 | (__v8hi)(__m128i)(B), (N))) |
||
| 294 | |||
| 295 | #define _mm_com_epu32(A, B, N) \ |
||
| 296 | ((__m128i)__builtin_ia32_vpcomud((__v4si)(__m128i)(A), \ |
||
| 297 | (__v4si)(__m128i)(B), (N))) |
||
| 298 | |||
| 299 | #define _mm_com_epu64(A, B, N) \ |
||
| 300 | ((__m128i)__builtin_ia32_vpcomuq((__v2di)(__m128i)(A), \ |
||
| 301 | (__v2di)(__m128i)(B), (N))) |
||
| 302 | |||
| 303 | #define _mm_com_epi8(A, B, N) \ |
||
| 304 | ((__m128i)__builtin_ia32_vpcomb((__v16qi)(__m128i)(A), \ |
||
| 305 | (__v16qi)(__m128i)(B), (N))) |
||
| 306 | |||
| 307 | #define _mm_com_epi16(A, B, N) \ |
||
| 308 | ((__m128i)__builtin_ia32_vpcomw((__v8hi)(__m128i)(A), \ |
||
| 309 | (__v8hi)(__m128i)(B), (N))) |
||
| 310 | |||
| 311 | #define _mm_com_epi32(A, B, N) \ |
||
| 312 | ((__m128i)__builtin_ia32_vpcomd((__v4si)(__m128i)(A), \ |
||
| 313 | (__v4si)(__m128i)(B), (N))) |
||
| 314 | |||
| 315 | #define _mm_com_epi64(A, B, N) \ |
||
| 316 | ((__m128i)__builtin_ia32_vpcomq((__v2di)(__m128i)(A), \ |
||
| 317 | (__v2di)(__m128i)(B), (N))) |
||
| 318 | |||
| 319 | #define _MM_PCOMCTRL_LT 0 |
||
| 320 | #define _MM_PCOMCTRL_LE 1 |
||
| 321 | #define _MM_PCOMCTRL_GT 2 |
||
| 322 | #define _MM_PCOMCTRL_GE 3 |
||
| 323 | #define _MM_PCOMCTRL_EQ 4 |
||
| 324 | #define _MM_PCOMCTRL_NEQ 5 |
||
| 325 | #define _MM_PCOMCTRL_FALSE 6 |
||
| 326 | #define _MM_PCOMCTRL_TRUE 7 |
||
| 327 | |||
| 328 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 329 | _mm_comlt_epu8(__m128i __A, __m128i __B) |
||
| 330 | { |
||
| 331 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_LT); |
||
| 332 | } |
||
| 333 | |||
| 334 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 335 | _mm_comle_epu8(__m128i __A, __m128i __B) |
||
| 336 | { |
||
| 337 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_LE); |
||
| 338 | } |
||
| 339 | |||
| 340 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 341 | _mm_comgt_epu8(__m128i __A, __m128i __B) |
||
| 342 | { |
||
| 343 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_GT); |
||
| 344 | } |
||
| 345 | |||
| 346 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 347 | _mm_comge_epu8(__m128i __A, __m128i __B) |
||
| 348 | { |
||
| 349 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_GE); |
||
| 350 | } |
||
| 351 | |||
| 352 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 353 | _mm_comeq_epu8(__m128i __A, __m128i __B) |
||
| 354 | { |
||
| 355 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 356 | } |
||
| 357 | |||
| 358 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 359 | _mm_comneq_epu8(__m128i __A, __m128i __B) |
||
| 360 | { |
||
| 361 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 362 | } |
||
| 363 | |||
| 364 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 365 | _mm_comfalse_epu8(__m128i __A, __m128i __B) |
||
| 366 | { |
||
| 367 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 368 | } |
||
| 369 | |||
| 370 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 371 | _mm_comtrue_epu8(__m128i __A, __m128i __B) |
||
| 372 | { |
||
| 373 | return _mm_com_epu8(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 374 | } |
||
| 375 | |||
| 376 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 377 | _mm_comlt_epu16(__m128i __A, __m128i __B) |
||
| 378 | { |
||
| 379 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_LT); |
||
| 380 | } |
||
| 381 | |||
| 382 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 383 | _mm_comle_epu16(__m128i __A, __m128i __B) |
||
| 384 | { |
||
| 385 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_LE); |
||
| 386 | } |
||
| 387 | |||
| 388 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 389 | _mm_comgt_epu16(__m128i __A, __m128i __B) |
||
| 390 | { |
||
| 391 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_GT); |
||
| 392 | } |
||
| 393 | |||
| 394 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 395 | _mm_comge_epu16(__m128i __A, __m128i __B) |
||
| 396 | { |
||
| 397 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_GE); |
||
| 398 | } |
||
| 399 | |||
| 400 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 401 | _mm_comeq_epu16(__m128i __A, __m128i __B) |
||
| 402 | { |
||
| 403 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 404 | } |
||
| 405 | |||
| 406 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 407 | _mm_comneq_epu16(__m128i __A, __m128i __B) |
||
| 408 | { |
||
| 409 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 410 | } |
||
| 411 | |||
| 412 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 413 | _mm_comfalse_epu16(__m128i __A, __m128i __B) |
||
| 414 | { |
||
| 415 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 416 | } |
||
| 417 | |||
| 418 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 419 | _mm_comtrue_epu16(__m128i __A, __m128i __B) |
||
| 420 | { |
||
| 421 | return _mm_com_epu16(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 422 | } |
||
| 423 | |||
| 424 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 425 | _mm_comlt_epu32(__m128i __A, __m128i __B) |
||
| 426 | { |
||
| 427 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_LT); |
||
| 428 | } |
||
| 429 | |||
| 430 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 431 | _mm_comle_epu32(__m128i __A, __m128i __B) |
||
| 432 | { |
||
| 433 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_LE); |
||
| 434 | } |
||
| 435 | |||
| 436 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 437 | _mm_comgt_epu32(__m128i __A, __m128i __B) |
||
| 438 | { |
||
| 439 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_GT); |
||
| 440 | } |
||
| 441 | |||
| 442 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 443 | _mm_comge_epu32(__m128i __A, __m128i __B) |
||
| 444 | { |
||
| 445 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_GE); |
||
| 446 | } |
||
| 447 | |||
| 448 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 449 | _mm_comeq_epu32(__m128i __A, __m128i __B) |
||
| 450 | { |
||
| 451 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 452 | } |
||
| 453 | |||
| 454 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 455 | _mm_comneq_epu32(__m128i __A, __m128i __B) |
||
| 456 | { |
||
| 457 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 458 | } |
||
| 459 | |||
| 460 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 461 | _mm_comfalse_epu32(__m128i __A, __m128i __B) |
||
| 462 | { |
||
| 463 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 464 | } |
||
| 465 | |||
| 466 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 467 | _mm_comtrue_epu32(__m128i __A, __m128i __B) |
||
| 468 | { |
||
| 469 | return _mm_com_epu32(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 470 | } |
||
| 471 | |||
| 472 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 473 | _mm_comlt_epu64(__m128i __A, __m128i __B) |
||
| 474 | { |
||
| 475 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_LT); |
||
| 476 | } |
||
| 477 | |||
| 478 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 479 | _mm_comle_epu64(__m128i __A, __m128i __B) |
||
| 480 | { |
||
| 481 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_LE); |
||
| 482 | } |
||
| 483 | |||
| 484 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 485 | _mm_comgt_epu64(__m128i __A, __m128i __B) |
||
| 486 | { |
||
| 487 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_GT); |
||
| 488 | } |
||
| 489 | |||
| 490 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 491 | _mm_comge_epu64(__m128i __A, __m128i __B) |
||
| 492 | { |
||
| 493 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_GE); |
||
| 494 | } |
||
| 495 | |||
| 496 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 497 | _mm_comeq_epu64(__m128i __A, __m128i __B) |
||
| 498 | { |
||
| 499 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 500 | } |
||
| 501 | |||
| 502 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 503 | _mm_comneq_epu64(__m128i __A, __m128i __B) |
||
| 504 | { |
||
| 505 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 506 | } |
||
| 507 | |||
| 508 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 509 | _mm_comfalse_epu64(__m128i __A, __m128i __B) |
||
| 510 | { |
||
| 511 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 512 | } |
||
| 513 | |||
| 514 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 515 | _mm_comtrue_epu64(__m128i __A, __m128i __B) |
||
| 516 | { |
||
| 517 | return _mm_com_epu64(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 518 | } |
||
| 519 | |||
| 520 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 521 | _mm_comlt_epi8(__m128i __A, __m128i __B) |
||
| 522 | { |
||
| 523 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_LT); |
||
| 524 | } |
||
| 525 | |||
| 526 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 527 | _mm_comle_epi8(__m128i __A, __m128i __B) |
||
| 528 | { |
||
| 529 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_LE); |
||
| 530 | } |
||
| 531 | |||
| 532 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 533 | _mm_comgt_epi8(__m128i __A, __m128i __B) |
||
| 534 | { |
||
| 535 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_GT); |
||
| 536 | } |
||
| 537 | |||
| 538 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 539 | _mm_comge_epi8(__m128i __A, __m128i __B) |
||
| 540 | { |
||
| 541 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_GE); |
||
| 542 | } |
||
| 543 | |||
| 544 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 545 | _mm_comeq_epi8(__m128i __A, __m128i __B) |
||
| 546 | { |
||
| 547 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 548 | } |
||
| 549 | |||
| 550 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 551 | _mm_comneq_epi8(__m128i __A, __m128i __B) |
||
| 552 | { |
||
| 553 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 554 | } |
||
| 555 | |||
| 556 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 557 | _mm_comfalse_epi8(__m128i __A, __m128i __B) |
||
| 558 | { |
||
| 559 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 560 | } |
||
| 561 | |||
| 562 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 563 | _mm_comtrue_epi8(__m128i __A, __m128i __B) |
||
| 564 | { |
||
| 565 | return _mm_com_epi8(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 566 | } |
||
| 567 | |||
| 568 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 569 | _mm_comlt_epi16(__m128i __A, __m128i __B) |
||
| 570 | { |
||
| 571 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_LT); |
||
| 572 | } |
||
| 573 | |||
| 574 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 575 | _mm_comle_epi16(__m128i __A, __m128i __B) |
||
| 576 | { |
||
| 577 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_LE); |
||
| 578 | } |
||
| 579 | |||
| 580 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 581 | _mm_comgt_epi16(__m128i __A, __m128i __B) |
||
| 582 | { |
||
| 583 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_GT); |
||
| 584 | } |
||
| 585 | |||
| 586 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 587 | _mm_comge_epi16(__m128i __A, __m128i __B) |
||
| 588 | { |
||
| 589 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_GE); |
||
| 590 | } |
||
| 591 | |||
| 592 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 593 | _mm_comeq_epi16(__m128i __A, __m128i __B) |
||
| 594 | { |
||
| 595 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 596 | } |
||
| 597 | |||
| 598 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 599 | _mm_comneq_epi16(__m128i __A, __m128i __B) |
||
| 600 | { |
||
| 601 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 602 | } |
||
| 603 | |||
| 604 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 605 | _mm_comfalse_epi16(__m128i __A, __m128i __B) |
||
| 606 | { |
||
| 607 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 608 | } |
||
| 609 | |||
| 610 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 611 | _mm_comtrue_epi16(__m128i __A, __m128i __B) |
||
| 612 | { |
||
| 613 | return _mm_com_epi16(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 614 | } |
||
| 615 | |||
| 616 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 617 | _mm_comlt_epi32(__m128i __A, __m128i __B) |
||
| 618 | { |
||
| 619 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_LT); |
||
| 620 | } |
||
| 621 | |||
| 622 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 623 | _mm_comle_epi32(__m128i __A, __m128i __B) |
||
| 624 | { |
||
| 625 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_LE); |
||
| 626 | } |
||
| 627 | |||
| 628 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 629 | _mm_comgt_epi32(__m128i __A, __m128i __B) |
||
| 630 | { |
||
| 631 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_GT); |
||
| 632 | } |
||
| 633 | |||
| 634 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 635 | _mm_comge_epi32(__m128i __A, __m128i __B) |
||
| 636 | { |
||
| 637 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_GE); |
||
| 638 | } |
||
| 639 | |||
| 640 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 641 | _mm_comeq_epi32(__m128i __A, __m128i __B) |
||
| 642 | { |
||
| 643 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 644 | } |
||
| 645 | |||
| 646 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 647 | _mm_comneq_epi32(__m128i __A, __m128i __B) |
||
| 648 | { |
||
| 649 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 650 | } |
||
| 651 | |||
| 652 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 653 | _mm_comfalse_epi32(__m128i __A, __m128i __B) |
||
| 654 | { |
||
| 655 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 656 | } |
||
| 657 | |||
| 658 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 659 | _mm_comtrue_epi32(__m128i __A, __m128i __B) |
||
| 660 | { |
||
| 661 | return _mm_com_epi32(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 662 | } |
||
| 663 | |||
| 664 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 665 | _mm_comlt_epi64(__m128i __A, __m128i __B) |
||
| 666 | { |
||
| 667 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_LT); |
||
| 668 | } |
||
| 669 | |||
| 670 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 671 | _mm_comle_epi64(__m128i __A, __m128i __B) |
||
| 672 | { |
||
| 673 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_LE); |
||
| 674 | } |
||
| 675 | |||
| 676 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 677 | _mm_comgt_epi64(__m128i __A, __m128i __B) |
||
| 678 | { |
||
| 679 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_GT); |
||
| 680 | } |
||
| 681 | |||
| 682 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 683 | _mm_comge_epi64(__m128i __A, __m128i __B) |
||
| 684 | { |
||
| 685 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_GE); |
||
| 686 | } |
||
| 687 | |||
| 688 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 689 | _mm_comeq_epi64(__m128i __A, __m128i __B) |
||
| 690 | { |
||
| 691 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_EQ); |
||
| 692 | } |
||
| 693 | |||
| 694 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 695 | _mm_comneq_epi64(__m128i __A, __m128i __B) |
||
| 696 | { |
||
| 697 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_NEQ); |
||
| 698 | } |
||
| 699 | |||
| 700 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 701 | _mm_comfalse_epi64(__m128i __A, __m128i __B) |
||
| 702 | { |
||
| 703 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_FALSE); |
||
| 704 | } |
||
| 705 | |||
| 706 | static __inline__ __m128i __DEFAULT_FN_ATTRS |
||
| 707 | _mm_comtrue_epi64(__m128i __A, __m128i __B) |
||
| 708 | { |
||
| 709 | return _mm_com_epi64(__A, __B, _MM_PCOMCTRL_TRUE); |
||
| 710 | } |
||
| 711 | |||
| 712 | #define _mm_permute2_pd(X, Y, C, I) \ |
||
| 713 | ((__m128d)__builtin_ia32_vpermil2pd((__v2df)(__m128d)(X), \ |
||
| 714 | (__v2df)(__m128d)(Y), \ |
||
| 715 | (__v2di)(__m128i)(C), (I))) |
||
| 716 | |||
| 717 | #define _mm256_permute2_pd(X, Y, C, I) \ |
||
| 718 | ((__m256d)__builtin_ia32_vpermil2pd256((__v4df)(__m256d)(X), \ |
||
| 719 | (__v4df)(__m256d)(Y), \ |
||
| 720 | (__v4di)(__m256i)(C), (I))) |
||
| 721 | |||
| 722 | #define _mm_permute2_ps(X, Y, C, I) \ |
||
| 723 | ((__m128)__builtin_ia32_vpermil2ps((__v4sf)(__m128)(X), (__v4sf)(__m128)(Y), \ |
||
| 724 | (__v4si)(__m128i)(C), (I))) |
||
| 725 | |||
| 726 | #define _mm256_permute2_ps(X, Y, C, I) \ |
||
| 727 | ((__m256)__builtin_ia32_vpermil2ps256((__v8sf)(__m256)(X), \ |
||
| 728 | (__v8sf)(__m256)(Y), \ |
||
| 729 | (__v8si)(__m256i)(C), (I))) |
||
| 730 | |||
| 731 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
||
| 732 | _mm_frcz_ss(__m128 __A) |
||
| 733 | { |
||
| 734 | return (__m128)__builtin_ia32_vfrczss((__v4sf)__A); |
||
| 735 | } |
||
| 736 | |||
| 737 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
||
| 738 | _mm_frcz_sd(__m128d __A) |
||
| 739 | { |
||
| 740 | return (__m128d)__builtin_ia32_vfrczsd((__v2df)__A); |
||
| 741 | } |
||
| 742 | |||
| 743 | static __inline__ __m128 __DEFAULT_FN_ATTRS |
||
| 744 | _mm_frcz_ps(__m128 __A) |
||
| 745 | { |
||
| 746 | return (__m128)__builtin_ia32_vfrczps((__v4sf)__A); |
||
| 747 | } |
||
| 748 | |||
| 749 | static __inline__ __m128d __DEFAULT_FN_ATTRS |
||
| 750 | _mm_frcz_pd(__m128d __A) |
||
| 751 | { |
||
| 752 | return (__m128d)__builtin_ia32_vfrczpd((__v2df)__A); |
||
| 753 | } |
||
| 754 | |||
| 755 | static __inline__ __m256 __DEFAULT_FN_ATTRS256 |
||
| 756 | _mm256_frcz_ps(__m256 __A) |
||
| 757 | { |
||
| 758 | return (__m256)__builtin_ia32_vfrczps256((__v8sf)__A); |
||
| 759 | } |
||
| 760 | |||
| 761 | static __inline__ __m256d __DEFAULT_FN_ATTRS256 |
||
| 762 | _mm256_frcz_pd(__m256d __A) |
||
| 763 | { |
||
| 764 | return (__m256d)__builtin_ia32_vfrczpd256((__v4df)__A); |
||
| 765 | } |
||
| 766 | |||
| 767 | #undef __DEFAULT_FN_ATTRS |
||
| 768 | #undef __DEFAULT_FN_ATTRS256 |
||
| 769 | |||
| 770 | #endif /* __XOPINTRIN_H */ |