OLD | NEW |
(Empty) | |
| 1 /*===---- emmintrin.h - SSE2 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 __EMMINTRIN_H |
| 25 #define __EMMINTRIN_H |
| 26 |
| 27 #ifndef __SSE2__ |
| 28 #error "SSE2 instruction set not enabled" |
| 29 #else |
| 30 |
| 31 #include <xmmintrin.h> |
| 32 |
| 33 typedef double __m128d __attribute__((__vector_size__(16))); |
| 34 typedef long long __m128i __attribute__((__vector_size__(16))); |
| 35 |
| 36 /* Type defines. */ |
| 37 typedef double __v2df __attribute__ ((__vector_size__ (16))); |
| 38 typedef long long __v2di __attribute__ ((__vector_size__ (16))); |
| 39 typedef short __v8hi __attribute__((__vector_size__(16))); |
| 40 typedef char __v16qi __attribute__((__vector_size__(16))); |
| 41 |
| 42 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 43 _mm_add_sd(__m128d a, __m128d b) |
| 44 { |
| 45 a[0] += b[0]; |
| 46 return a; |
| 47 } |
| 48 |
| 49 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 50 _mm_add_pd(__m128d a, __m128d b) |
| 51 { |
| 52 return a + b; |
| 53 } |
| 54 |
| 55 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 56 _mm_sub_sd(__m128d a, __m128d b) |
| 57 { |
| 58 a[0] -= b[0]; |
| 59 return a; |
| 60 } |
| 61 |
| 62 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 63 _mm_sub_pd(__m128d a, __m128d b) |
| 64 { |
| 65 return a - b; |
| 66 } |
| 67 |
| 68 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 69 _mm_mul_sd(__m128d a, __m128d b) |
| 70 { |
| 71 a[0] *= b[0]; |
| 72 return a; |
| 73 } |
| 74 |
| 75 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 76 _mm_mul_pd(__m128d a, __m128d b) |
| 77 { |
| 78 return a * b; |
| 79 } |
| 80 |
| 81 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 82 _mm_div_sd(__m128d a, __m128d b) |
| 83 { |
| 84 a[0] /= b[0]; |
| 85 return a; |
| 86 } |
| 87 |
| 88 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 89 _mm_div_pd(__m128d a, __m128d b) |
| 90 { |
| 91 return a / b; |
| 92 } |
| 93 |
| 94 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 95 _mm_sqrt_sd(__m128d a, __m128d b) |
| 96 { |
| 97 __m128d c = __builtin_ia32_sqrtsd(b); |
| 98 return (__m128d) { c[0], a[1] }; |
| 99 } |
| 100 |
| 101 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 102 _mm_sqrt_pd(__m128d a) |
| 103 { |
| 104 return __builtin_ia32_sqrtpd(a); |
| 105 } |
| 106 |
| 107 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 108 _mm_min_sd(__m128d a, __m128d b) |
| 109 { |
| 110 return __builtin_ia32_minsd(a, b); |
| 111 } |
| 112 |
| 113 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 114 _mm_min_pd(__m128d a, __m128d b) |
| 115 { |
| 116 return __builtin_ia32_minpd(a, b); |
| 117 } |
| 118 |
| 119 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 120 _mm_max_sd(__m128d a, __m128d b) |
| 121 { |
| 122 return __builtin_ia32_maxsd(a, b); |
| 123 } |
| 124 |
| 125 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 126 _mm_max_pd(__m128d a, __m128d b) |
| 127 { |
| 128 return __builtin_ia32_maxpd(a, b); |
| 129 } |
| 130 |
| 131 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 132 _mm_and_pd(__m128d a, __m128d b) |
| 133 { |
| 134 return (__m128d)((__v4si)a & (__v4si)b); |
| 135 } |
| 136 |
| 137 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 138 _mm_andnot_pd(__m128d a, __m128d b) |
| 139 { |
| 140 return (__m128d)(~(__v4si)a & (__v4si)b); |
| 141 } |
| 142 |
| 143 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 144 _mm_or_pd(__m128d a, __m128d b) |
| 145 { |
| 146 return (__m128d)((__v4si)a | (__v4si)b); |
| 147 } |
| 148 |
| 149 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 150 _mm_xor_pd(__m128d a, __m128d b) |
| 151 { |
| 152 return (__m128d)((__v4si)a ^ (__v4si)b); |
| 153 } |
| 154 |
| 155 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 156 _mm_cmpeq_pd(__m128d a, __m128d b) |
| 157 { |
| 158 return (__m128d)__builtin_ia32_cmppd(a, b, 0); |
| 159 } |
| 160 |
| 161 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 162 _mm_cmplt_pd(__m128d a, __m128d b) |
| 163 { |
| 164 return (__m128d)__builtin_ia32_cmppd(a, b, 1); |
| 165 } |
| 166 |
| 167 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 168 _mm_cmple_pd(__m128d a, __m128d b) |
| 169 { |
| 170 return (__m128d)__builtin_ia32_cmppd(a, b, 2); |
| 171 } |
| 172 |
| 173 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 174 _mm_cmpgt_pd(__m128d a, __m128d b) |
| 175 { |
| 176 return (__m128d)__builtin_ia32_cmppd(b, a, 1); |
| 177 } |
| 178 |
| 179 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 180 _mm_cmpge_pd(__m128d a, __m128d b) |
| 181 { |
| 182 return (__m128d)__builtin_ia32_cmppd(b, a, 2); |
| 183 } |
| 184 |
| 185 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 186 _mm_cmpord_pd(__m128d a, __m128d b) |
| 187 { |
| 188 return (__m128d)__builtin_ia32_cmppd(a, b, 7); |
| 189 } |
| 190 |
| 191 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 192 _mm_cmpunord_pd(__m128d a, __m128d b) |
| 193 { |
| 194 return (__m128d)__builtin_ia32_cmppd(a, b, 3); |
| 195 } |
| 196 |
| 197 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 198 _mm_cmpneq_pd(__m128d a, __m128d b) |
| 199 { |
| 200 return (__m128d)__builtin_ia32_cmppd(a, b, 4); |
| 201 } |
| 202 |
| 203 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 204 _mm_cmpnlt_pd(__m128d a, __m128d b) |
| 205 { |
| 206 return (__m128d)__builtin_ia32_cmppd(a, b, 5); |
| 207 } |
| 208 |
| 209 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 210 _mm_cmpnle_pd(__m128d a, __m128d b) |
| 211 { |
| 212 return (__m128d)__builtin_ia32_cmppd(a, b, 6); |
| 213 } |
| 214 |
| 215 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 216 _mm_cmpngt_pd(__m128d a, __m128d b) |
| 217 { |
| 218 return (__m128d)__builtin_ia32_cmppd(b, a, 5); |
| 219 } |
| 220 |
| 221 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 222 _mm_cmpnge_pd(__m128d a, __m128d b) |
| 223 { |
| 224 return (__m128d)__builtin_ia32_cmppd(b, a, 6); |
| 225 } |
| 226 |
| 227 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 228 _mm_cmpeq_sd(__m128d a, __m128d b) |
| 229 { |
| 230 return (__m128d)__builtin_ia32_cmpsd(a, b, 0); |
| 231 } |
| 232 |
| 233 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 234 _mm_cmplt_sd(__m128d a, __m128d b) |
| 235 { |
| 236 return (__m128d)__builtin_ia32_cmpsd(a, b, 1); |
| 237 } |
| 238 |
| 239 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 240 _mm_cmple_sd(__m128d a, __m128d b) |
| 241 { |
| 242 return (__m128d)__builtin_ia32_cmpsd(a, b, 2); |
| 243 } |
| 244 |
| 245 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 246 _mm_cmpgt_sd(__m128d a, __m128d b) |
| 247 { |
| 248 return (__m128d)__builtin_ia32_cmpsd(b, a, 1); |
| 249 } |
| 250 |
| 251 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 252 _mm_cmpge_sd(__m128d a, __m128d b) |
| 253 { |
| 254 return (__m128d)__builtin_ia32_cmpsd(b, a, 2); |
| 255 } |
| 256 |
| 257 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 258 _mm_cmpord_sd(__m128d a, __m128d b) |
| 259 { |
| 260 return (__m128d)__builtin_ia32_cmpsd(a, b, 7); |
| 261 } |
| 262 |
| 263 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 264 _mm_cmpunord_sd(__m128d a, __m128d b) |
| 265 { |
| 266 return (__m128d)__builtin_ia32_cmpsd(a, b, 3); |
| 267 } |
| 268 |
| 269 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 270 _mm_cmpneq_sd(__m128d a, __m128d b) |
| 271 { |
| 272 return (__m128d)__builtin_ia32_cmpsd(a, b, 4); |
| 273 } |
| 274 |
| 275 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 276 _mm_cmpnlt_sd(__m128d a, __m128d b) |
| 277 { |
| 278 return (__m128d)__builtin_ia32_cmpsd(a, b, 5); |
| 279 } |
| 280 |
| 281 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 282 _mm_cmpnle_sd(__m128d a, __m128d b) |
| 283 { |
| 284 return (__m128d)__builtin_ia32_cmpsd(a, b, 6); |
| 285 } |
| 286 |
| 287 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 288 _mm_cmpngt_sd(__m128d a, __m128d b) |
| 289 { |
| 290 return (__m128d)__builtin_ia32_cmpsd(b, a, 5); |
| 291 } |
| 292 |
| 293 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 294 _mm_cmpnge_sd(__m128d a, __m128d b) |
| 295 { |
| 296 return (__m128d)__builtin_ia32_cmpsd(b, a, 6); |
| 297 } |
| 298 |
| 299 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 300 _mm_comieq_sd(__m128d a, __m128d b) |
| 301 { |
| 302 return __builtin_ia32_comisdeq(a, b); |
| 303 } |
| 304 |
| 305 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 306 _mm_comilt_sd(__m128d a, __m128d b) |
| 307 { |
| 308 return __builtin_ia32_comisdlt(a, b); |
| 309 } |
| 310 |
| 311 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 312 _mm_comile_sd(__m128d a, __m128d b) |
| 313 { |
| 314 return __builtin_ia32_comisdle(a, b); |
| 315 } |
| 316 |
| 317 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 318 _mm_comigt_sd(__m128d a, __m128d b) |
| 319 { |
| 320 return __builtin_ia32_comisdgt(a, b); |
| 321 } |
| 322 |
| 323 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 324 _mm_comige_sd(__m128d a, __m128d b) |
| 325 { |
| 326 return __builtin_ia32_comisdge(a, b); |
| 327 } |
| 328 |
| 329 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 330 _mm_comineq_sd(__m128d a, __m128d b) |
| 331 { |
| 332 return __builtin_ia32_comisdneq(a, b); |
| 333 } |
| 334 |
| 335 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 336 _mm_ucomieq_sd(__m128d a, __m128d b) |
| 337 { |
| 338 return __builtin_ia32_ucomisdeq(a, b); |
| 339 } |
| 340 |
| 341 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 342 _mm_ucomilt_sd(__m128d a, __m128d b) |
| 343 { |
| 344 return __builtin_ia32_ucomisdlt(a, b); |
| 345 } |
| 346 |
| 347 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 348 _mm_ucomile_sd(__m128d a, __m128d b) |
| 349 { |
| 350 return __builtin_ia32_ucomisdle(a, b); |
| 351 } |
| 352 |
| 353 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 354 _mm_ucomigt_sd(__m128d a, __m128d b) |
| 355 { |
| 356 return __builtin_ia32_ucomisdgt(a, b); |
| 357 } |
| 358 |
| 359 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 360 _mm_ucomige_sd(__m128d a, __m128d b) |
| 361 { |
| 362 return __builtin_ia32_ucomisdge(a, b); |
| 363 } |
| 364 |
| 365 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 366 _mm_ucomineq_sd(__m128d a, __m128d b) |
| 367 { |
| 368 return __builtin_ia32_ucomisdneq(a, b); |
| 369 } |
| 370 |
| 371 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 372 _mm_cvtpd_ps(__m128d a) |
| 373 { |
| 374 return __builtin_ia32_cvtpd2ps(a); |
| 375 } |
| 376 |
| 377 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 378 _mm_cvtps_pd(__m128 a) |
| 379 { |
| 380 return __builtin_ia32_cvtps2pd(a); |
| 381 } |
| 382 |
| 383 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 384 _mm_cvtepi32_pd(__m128i a) |
| 385 { |
| 386 return __builtin_ia32_cvtdq2pd((__v4si)a); |
| 387 } |
| 388 |
| 389 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 390 _mm_cvtpd_epi32(__m128d a) |
| 391 { |
| 392 return __builtin_ia32_cvtpd2dq(a); |
| 393 } |
| 394 |
| 395 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 396 _mm_cvtsd_si32(__m128d a) |
| 397 { |
| 398 return __builtin_ia32_cvtsd2si(a); |
| 399 } |
| 400 |
| 401 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 402 _mm_cvtsd_ss(__m128 a, __m128d b) |
| 403 { |
| 404 a[0] = b[0]; |
| 405 return a; |
| 406 } |
| 407 |
| 408 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 409 _mm_cvtsi32_sd(__m128d a, int b) |
| 410 { |
| 411 a[0] = b; |
| 412 return a; |
| 413 } |
| 414 |
| 415 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 416 _mm_cvtss_sd(__m128d a, __m128 b) |
| 417 { |
| 418 a[0] = b[0]; |
| 419 return a; |
| 420 } |
| 421 |
| 422 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 423 _mm_cvttpd_epi32(__m128d a) |
| 424 { |
| 425 return (__m128i)__builtin_ia32_cvttpd2dq(a); |
| 426 } |
| 427 |
| 428 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 429 _mm_cvttsd_si32(__m128d a) |
| 430 { |
| 431 return a[0]; |
| 432 } |
| 433 |
| 434 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) |
| 435 _mm_cvtpd_pi32(__m128d a) |
| 436 { |
| 437 return (__m64)__builtin_ia32_cvtpd2pi(a); |
| 438 } |
| 439 |
| 440 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) |
| 441 _mm_cvttpd_pi32(__m128d a) |
| 442 { |
| 443 return (__m64)__builtin_ia32_cvttpd2pi(a); |
| 444 } |
| 445 |
| 446 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 447 _mm_cvtpi32_pd(__m64 a) |
| 448 { |
| 449 return __builtin_ia32_cvtpi2pd((__v2si)a); |
| 450 } |
| 451 |
| 452 static __inline__ double __attribute__((__always_inline__, __nodebug__)) |
| 453 _mm_cvtsd_f64(__m128d a) |
| 454 { |
| 455 return a[0]; |
| 456 } |
| 457 |
| 458 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 459 _mm_load_pd(double const *dp) |
| 460 { |
| 461 return *(__m128d*)dp; |
| 462 } |
| 463 |
| 464 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 465 _mm_load1_pd(double const *dp) |
| 466 { |
| 467 struct __mm_load1_pd_struct { |
| 468 double u; |
| 469 } __attribute__((__packed__, __may_alias__)); |
| 470 double u = ((struct __mm_load1_pd_struct*)dp)->u; |
| 471 return (__m128d){ u, u }; |
| 472 } |
| 473 |
| 474 #define _mm_load_pd1(dp) _mm_load1_pd(dp) |
| 475 |
| 476 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 477 _mm_loadr_pd(double const *dp) |
| 478 { |
| 479 __m128d u = *(__m128d*)dp; |
| 480 return __builtin_shufflevector(u, u, 1, 0); |
| 481 } |
| 482 |
| 483 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 484 _mm_loadu_pd(double const *dp) |
| 485 { |
| 486 struct __loadu_pd { |
| 487 __m128d v; |
| 488 } __attribute__((packed, may_alias)); |
| 489 return ((struct __loadu_pd*)dp)->v; |
| 490 } |
| 491 |
| 492 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 493 _mm_load_sd(double const *dp) |
| 494 { |
| 495 struct __mm_load_sd_struct { |
| 496 double u; |
| 497 } __attribute__((__packed__, __may_alias__)); |
| 498 double u = ((struct __mm_load_sd_struct*)dp)->u; |
| 499 return (__m128d){ u, 0 }; |
| 500 } |
| 501 |
| 502 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 503 _mm_loadh_pd(__m128d a, double const *dp) |
| 504 { |
| 505 struct __mm_loadh_pd_struct { |
| 506 double u; |
| 507 } __attribute__((__packed__, __may_alias__)); |
| 508 double u = ((struct __mm_loadh_pd_struct*)dp)->u; |
| 509 return (__m128d){ a[0], u }; |
| 510 } |
| 511 |
| 512 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 513 _mm_loadl_pd(__m128d a, double const *dp) |
| 514 { |
| 515 struct __mm_loadl_pd_struct { |
| 516 double u; |
| 517 } __attribute__((__packed__, __may_alias__)); |
| 518 double u = ((struct __mm_loadl_pd_struct*)dp)->u; |
| 519 return (__m128d){ u, a[1] }; |
| 520 } |
| 521 |
| 522 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 523 _mm_set_sd(double w) |
| 524 { |
| 525 return (__m128d){ w, 0 }; |
| 526 } |
| 527 |
| 528 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 529 _mm_set1_pd(double w) |
| 530 { |
| 531 return (__m128d){ w, w }; |
| 532 } |
| 533 |
| 534 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 535 _mm_set_pd(double w, double x) |
| 536 { |
| 537 return (__m128d){ x, w }; |
| 538 } |
| 539 |
| 540 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 541 _mm_setr_pd(double w, double x) |
| 542 { |
| 543 return (__m128d){ w, x }; |
| 544 } |
| 545 |
| 546 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 547 _mm_setzero_pd(void) |
| 548 { |
| 549 return (__m128d){ 0, 0 }; |
| 550 } |
| 551 |
| 552 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 553 _mm_move_sd(__m128d a, __m128d b) |
| 554 { |
| 555 return (__m128d){ b[0], a[1] }; |
| 556 } |
| 557 |
| 558 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 559 _mm_store_sd(double *dp, __m128d a) |
| 560 { |
| 561 struct __mm_store_sd_struct { |
| 562 double u; |
| 563 } __attribute__((__packed__, __may_alias__)); |
| 564 ((struct __mm_store_sd_struct*)dp)->u = a[0]; |
| 565 } |
| 566 |
| 567 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 568 _mm_store1_pd(double *dp, __m128d a) |
| 569 { |
| 570 struct __mm_store1_pd_struct { |
| 571 double u[2]; |
| 572 } __attribute__((__packed__, __may_alias__)); |
| 573 ((struct __mm_store1_pd_struct*)dp)->u[0] = a[0]; |
| 574 ((struct __mm_store1_pd_struct*)dp)->u[1] = a[0]; |
| 575 } |
| 576 |
| 577 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 578 _mm_store_pd(double *dp, __m128d a) |
| 579 { |
| 580 *(__m128d *)dp = a; |
| 581 } |
| 582 |
| 583 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 584 _mm_storeu_pd(double *dp, __m128d a) |
| 585 { |
| 586 __builtin_ia32_storeupd(dp, a); |
| 587 } |
| 588 |
| 589 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 590 _mm_storer_pd(double *dp, __m128d a) |
| 591 { |
| 592 a = __builtin_shufflevector(a, a, 1, 0); |
| 593 *(__m128d *)dp = a; |
| 594 } |
| 595 |
| 596 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 597 _mm_storeh_pd(double *dp, __m128d a) |
| 598 { |
| 599 struct __mm_storeh_pd_struct { |
| 600 double u; |
| 601 } __attribute__((__packed__, __may_alias__)); |
| 602 ((struct __mm_storeh_pd_struct*)dp)->u = a[1]; |
| 603 } |
| 604 |
| 605 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 606 _mm_storel_pd(double *dp, __m128d a) |
| 607 { |
| 608 struct __mm_storeh_pd_struct { |
| 609 double u; |
| 610 } __attribute__((__packed__, __may_alias__)); |
| 611 ((struct __mm_storeh_pd_struct*)dp)->u = a[0]; |
| 612 } |
| 613 |
| 614 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 615 _mm_add_epi8(__m128i a, __m128i b) |
| 616 { |
| 617 return (__m128i)((__v16qi)a + (__v16qi)b); |
| 618 } |
| 619 |
| 620 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 621 _mm_add_epi16(__m128i a, __m128i b) |
| 622 { |
| 623 return (__m128i)((__v8hi)a + (__v8hi)b); |
| 624 } |
| 625 |
| 626 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 627 _mm_add_epi32(__m128i a, __m128i b) |
| 628 { |
| 629 return (__m128i)((__v4si)a + (__v4si)b); |
| 630 } |
| 631 |
| 632 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) |
| 633 _mm_add_si64(__m64 a, __m64 b) |
| 634 { |
| 635 return a + b; |
| 636 } |
| 637 |
| 638 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 639 _mm_add_epi64(__m128i a, __m128i b) |
| 640 { |
| 641 return a + b; |
| 642 } |
| 643 |
| 644 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 645 _mm_adds_epi8(__m128i a, __m128i b) |
| 646 { |
| 647 return (__m128i)__builtin_ia32_paddsb128((__v16qi)a, (__v16qi)b); |
| 648 } |
| 649 |
| 650 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 651 _mm_adds_epi16(__m128i a, __m128i b) |
| 652 { |
| 653 return (__m128i)__builtin_ia32_paddsw128((__v8hi)a, (__v8hi)b); |
| 654 } |
| 655 |
| 656 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 657 _mm_adds_epu8(__m128i a, __m128i b) |
| 658 { |
| 659 return (__m128i)__builtin_ia32_paddusb128((__v16qi)a, (__v16qi)b); |
| 660 } |
| 661 |
| 662 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 663 _mm_adds_epu16(__m128i a, __m128i b) |
| 664 { |
| 665 return (__m128i)__builtin_ia32_paddusw128((__v8hi)a, (__v8hi)b); |
| 666 } |
| 667 |
| 668 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 669 _mm_avg_epu8(__m128i a, __m128i b) |
| 670 { |
| 671 return (__m128i)__builtin_ia32_pavgb128((__v16qi)a, (__v16qi)b); |
| 672 } |
| 673 |
| 674 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 675 _mm_avg_epu16(__m128i a, __m128i b) |
| 676 { |
| 677 return (__m128i)__builtin_ia32_pavgw128((__v8hi)a, (__v8hi)b); |
| 678 } |
| 679 |
| 680 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 681 _mm_madd_epi16(__m128i a, __m128i b) |
| 682 { |
| 683 return (__m128i)__builtin_ia32_pmaddwd128((__v8hi)a, (__v8hi)b); |
| 684 } |
| 685 |
| 686 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 687 _mm_max_epi16(__m128i a, __m128i b) |
| 688 { |
| 689 return (__m128i)__builtin_ia32_pmaxsw128((__v8hi)a, (__v8hi)b); |
| 690 } |
| 691 |
| 692 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 693 _mm_max_epu8(__m128i a, __m128i b) |
| 694 { |
| 695 return (__m128i)__builtin_ia32_pmaxub128((__v16qi)a, (__v16qi)b); |
| 696 } |
| 697 |
| 698 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 699 _mm_min_epi16(__m128i a, __m128i b) |
| 700 { |
| 701 return (__m128i)__builtin_ia32_pminsw128((__v8hi)a, (__v8hi)b); |
| 702 } |
| 703 |
| 704 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 705 _mm_min_epu8(__m128i a, __m128i b) |
| 706 { |
| 707 return (__m128i)__builtin_ia32_pminub128((__v16qi)a, (__v16qi)b); |
| 708 } |
| 709 |
| 710 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 711 _mm_mulhi_epi16(__m128i a, __m128i b) |
| 712 { |
| 713 return (__m128i)__builtin_ia32_pmulhw128((__v8hi)a, (__v8hi)b); |
| 714 } |
| 715 |
| 716 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 717 _mm_mulhi_epu16(__m128i a, __m128i b) |
| 718 { |
| 719 return (__m128i)__builtin_ia32_pmulhuw128((__v8hi)a, (__v8hi)b); |
| 720 } |
| 721 |
| 722 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 723 _mm_mullo_epi16(__m128i a, __m128i b) |
| 724 { |
| 725 return (__m128i)((__v8hi)a * (__v8hi)b); |
| 726 } |
| 727 |
| 728 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) |
| 729 _mm_mul_su32(__m64 a, __m64 b) |
| 730 { |
| 731 return __builtin_ia32_pmuludq((__v2si)a, (__v2si)b); |
| 732 } |
| 733 |
| 734 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 735 _mm_mul_epu32(__m128i a, __m128i b) |
| 736 { |
| 737 return __builtin_ia32_pmuludq128((__v4si)a, (__v4si)b); |
| 738 } |
| 739 |
| 740 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 741 _mm_sad_epu8(__m128i a, __m128i b) |
| 742 { |
| 743 return __builtin_ia32_psadbw128((__v16qi)a, (__v16qi)b); |
| 744 } |
| 745 |
| 746 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 747 _mm_sub_epi8(__m128i a, __m128i b) |
| 748 { |
| 749 return (__m128i)((__v16qi)a - (__v16qi)b); |
| 750 } |
| 751 |
| 752 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 753 _mm_sub_epi16(__m128i a, __m128i b) |
| 754 { |
| 755 return (__m128i)((__v8hi)a - (__v8hi)b); |
| 756 } |
| 757 |
| 758 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 759 _mm_sub_epi32(__m128i a, __m128i b) |
| 760 { |
| 761 return (__m128i)((__v4si)a - (__v4si)b); |
| 762 } |
| 763 |
| 764 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) |
| 765 _mm_sub_si64(__m64 a, __m64 b) |
| 766 { |
| 767 return a - b; |
| 768 } |
| 769 |
| 770 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 771 _mm_sub_epi64(__m128i a, __m128i b) |
| 772 { |
| 773 return a - b; |
| 774 } |
| 775 |
| 776 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 777 _mm_subs_epi8(__m128i a, __m128i b) |
| 778 { |
| 779 return (__m128i)__builtin_ia32_psubsb128((__v16qi)a, (__v16qi)b); |
| 780 } |
| 781 |
| 782 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 783 _mm_subs_epi16(__m128i a, __m128i b) |
| 784 { |
| 785 return (__m128i)__builtin_ia32_psubsw128((__v8hi)a, (__v8hi)b); |
| 786 } |
| 787 |
| 788 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 789 _mm_subs_epu8(__m128i a, __m128i b) |
| 790 { |
| 791 return (__m128i)__builtin_ia32_psubusb128((__v16qi)a, (__v16qi)b); |
| 792 } |
| 793 |
| 794 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 795 _mm_subs_epu16(__m128i a, __m128i b) |
| 796 { |
| 797 return (__m128i)__builtin_ia32_psubusw128((__v8hi)a, (__v8hi)b); |
| 798 } |
| 799 |
| 800 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 801 _mm_and_si128(__m128i a, __m128i b) |
| 802 { |
| 803 return a & b; |
| 804 } |
| 805 |
| 806 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 807 _mm_andnot_si128(__m128i a, __m128i b) |
| 808 { |
| 809 return ~a & b; |
| 810 } |
| 811 |
| 812 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 813 _mm_or_si128(__m128i a, __m128i b) |
| 814 { |
| 815 return a | b; |
| 816 } |
| 817 |
| 818 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 819 _mm_xor_si128(__m128i a, __m128i b) |
| 820 { |
| 821 return a ^ b; |
| 822 } |
| 823 |
| 824 #define _mm_slli_si128(VEC, IMM) \ |
| 825 ((__m128i)__builtin_ia32_pslldqi128((__m128i)(VEC), (IMM)*8)) |
| 826 |
| 827 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 828 _mm_slli_epi16(__m128i a, int count) |
| 829 { |
| 830 return (__m128i)__builtin_ia32_psllwi128((__v8hi)a, count); |
| 831 } |
| 832 |
| 833 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 834 _mm_sll_epi16(__m128i a, __m128i count) |
| 835 { |
| 836 return (__m128i)__builtin_ia32_psllw128((__v8hi)a, (__v8hi)count); |
| 837 } |
| 838 |
| 839 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 840 _mm_slli_epi32(__m128i a, int count) |
| 841 { |
| 842 return (__m128i)__builtin_ia32_pslldi128((__v4si)a, count); |
| 843 } |
| 844 |
| 845 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 846 _mm_sll_epi32(__m128i a, __m128i count) |
| 847 { |
| 848 return (__m128i)__builtin_ia32_pslld128((__v4si)a, (__v4si)count); |
| 849 } |
| 850 |
| 851 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 852 _mm_slli_epi64(__m128i a, int count) |
| 853 { |
| 854 return __builtin_ia32_psllqi128(a, count); |
| 855 } |
| 856 |
| 857 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 858 _mm_sll_epi64(__m128i a, __m128i count) |
| 859 { |
| 860 return __builtin_ia32_psllq128(a, count); |
| 861 } |
| 862 |
| 863 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 864 _mm_srai_epi16(__m128i a, int count) |
| 865 { |
| 866 return (__m128i)__builtin_ia32_psrawi128((__v8hi)a, count); |
| 867 } |
| 868 |
| 869 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 870 _mm_sra_epi16(__m128i a, __m128i count) |
| 871 { |
| 872 return (__m128i)__builtin_ia32_psraw128((__v8hi)a, (__v8hi)count); |
| 873 } |
| 874 |
| 875 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 876 _mm_srai_epi32(__m128i a, int count) |
| 877 { |
| 878 return (__m128i)__builtin_ia32_psradi128((__v4si)a, count); |
| 879 } |
| 880 |
| 881 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 882 _mm_sra_epi32(__m128i a, __m128i count) |
| 883 { |
| 884 return (__m128i)__builtin_ia32_psrad128((__v4si)a, (__v4si)count); |
| 885 } |
| 886 |
| 887 |
| 888 #define _mm_srli_si128(VEC, IMM) \ |
| 889 ((__m128i)__builtin_ia32_psrldqi128((__m128i)(VEC), (IMM)*8)) |
| 890 |
| 891 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 892 _mm_srli_epi16(__m128i a, int count) |
| 893 { |
| 894 return (__m128i)__builtin_ia32_psrlwi128((__v8hi)a, count); |
| 895 } |
| 896 |
| 897 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 898 _mm_srl_epi16(__m128i a, __m128i count) |
| 899 { |
| 900 return (__m128i)__builtin_ia32_psrlw128((__v8hi)a, (__v8hi)count); |
| 901 } |
| 902 |
| 903 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 904 _mm_srli_epi32(__m128i a, int count) |
| 905 { |
| 906 return (__m128i)__builtin_ia32_psrldi128((__v4si)a, count); |
| 907 } |
| 908 |
| 909 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 910 _mm_srl_epi32(__m128i a, __m128i count) |
| 911 { |
| 912 return (__m128i)__builtin_ia32_psrld128((__v4si)a, (__v4si)count); |
| 913 } |
| 914 |
| 915 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 916 _mm_srli_epi64(__m128i a, int count) |
| 917 { |
| 918 return __builtin_ia32_psrlqi128(a, count); |
| 919 } |
| 920 |
| 921 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 922 _mm_srl_epi64(__m128i a, __m128i count) |
| 923 { |
| 924 return __builtin_ia32_psrlq128(a, count); |
| 925 } |
| 926 |
| 927 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 928 _mm_cmpeq_epi8(__m128i a, __m128i b) |
| 929 { |
| 930 return (__m128i)((__v16qi)a == (__v16qi)b); |
| 931 } |
| 932 |
| 933 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 934 _mm_cmpeq_epi16(__m128i a, __m128i b) |
| 935 { |
| 936 return (__m128i)((__v8hi)a == (__v8hi)b); |
| 937 } |
| 938 |
| 939 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 940 _mm_cmpeq_epi32(__m128i a, __m128i b) |
| 941 { |
| 942 return (__m128i)((__v4si)a == (__v4si)b); |
| 943 } |
| 944 |
| 945 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 946 _mm_cmpgt_epi8(__m128i a, __m128i b) |
| 947 { |
| 948 return (__m128i)((__v16qi)a > (__v16qi)b); |
| 949 } |
| 950 |
| 951 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 952 _mm_cmpgt_epi16(__m128i a, __m128i b) |
| 953 { |
| 954 return (__m128i)((__v8hi)a > (__v8hi)b); |
| 955 } |
| 956 |
| 957 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 958 _mm_cmpgt_epi32(__m128i a, __m128i b) |
| 959 { |
| 960 return (__m128i)((__v4si)a > (__v4si)b); |
| 961 } |
| 962 |
| 963 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 964 _mm_cmplt_epi8(__m128i a, __m128i b) |
| 965 { |
| 966 return _mm_cmpgt_epi8(b,a); |
| 967 } |
| 968 |
| 969 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 970 _mm_cmplt_epi16(__m128i a, __m128i b) |
| 971 { |
| 972 return _mm_cmpgt_epi16(b,a); |
| 973 } |
| 974 |
| 975 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 976 _mm_cmplt_epi32(__m128i a, __m128i b) |
| 977 { |
| 978 return _mm_cmpgt_epi32(b,a); |
| 979 } |
| 980 |
| 981 #ifdef __x86_64__ |
| 982 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 983 _mm_cvtsi64_sd(__m128d a, long long b) |
| 984 { |
| 985 a[0] = b; |
| 986 return a; |
| 987 } |
| 988 |
| 989 static __inline__ long long __attribute__((__always_inline__, __nodebug__)) |
| 990 _mm_cvtsd_si64(__m128d a) |
| 991 { |
| 992 return __builtin_ia32_cvtsd2si64(a); |
| 993 } |
| 994 |
| 995 static __inline__ long long __attribute__((__always_inline__, __nodebug__)) |
| 996 _mm_cvttsd_si64(__m128d a) |
| 997 { |
| 998 return a[0]; |
| 999 } |
| 1000 #endif |
| 1001 |
| 1002 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 1003 _mm_cvtepi32_ps(__m128i a) |
| 1004 { |
| 1005 return __builtin_ia32_cvtdq2ps((__v4si)a); |
| 1006 } |
| 1007 |
| 1008 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1009 _mm_cvtps_epi32(__m128 a) |
| 1010 { |
| 1011 return (__m128i)__builtin_ia32_cvtps2dq(a); |
| 1012 } |
| 1013 |
| 1014 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1015 _mm_cvttps_epi32(__m128 a) |
| 1016 { |
| 1017 return (__m128i)__builtin_ia32_cvttps2dq(a); |
| 1018 } |
| 1019 |
| 1020 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1021 _mm_cvtsi32_si128(int a) |
| 1022 { |
| 1023 return (__m128i)(__v4si){ a, 0, 0, 0 }; |
| 1024 } |
| 1025 |
| 1026 #ifdef __x86_64__ |
| 1027 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1028 _mm_cvtsi64_si128(long long a) |
| 1029 { |
| 1030 return (__m128i){ a, 0 }; |
| 1031 } |
| 1032 #endif |
| 1033 |
| 1034 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 1035 _mm_cvtsi128_si32(__m128i a) |
| 1036 { |
| 1037 __v4si b = (__v4si)a; |
| 1038 return b[0]; |
| 1039 } |
| 1040 |
| 1041 #ifdef __x86_64__ |
| 1042 static __inline__ long long __attribute__((__always_inline__, __nodebug__)) |
| 1043 _mm_cvtsi128_si64(__m128i a) |
| 1044 { |
| 1045 return a[0]; |
| 1046 } |
| 1047 #endif |
| 1048 |
| 1049 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1050 _mm_load_si128(__m128i const *p) |
| 1051 { |
| 1052 return *p; |
| 1053 } |
| 1054 |
| 1055 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1056 _mm_loadu_si128(__m128i const *p) |
| 1057 { |
| 1058 struct __loadu_si128 { |
| 1059 __m128i v; |
| 1060 } __attribute__((packed, may_alias)); |
| 1061 return ((struct __loadu_si128*)p)->v; |
| 1062 } |
| 1063 |
| 1064 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1065 _mm_loadl_epi64(__m128i const *p) |
| 1066 { |
| 1067 struct __mm_loadl_epi64_struct { |
| 1068 long long u; |
| 1069 } __attribute__((__packed__, __may_alias__)); |
| 1070 return (__m128i) { ((struct __mm_loadl_epi64_struct*)p)->u, 0}; |
| 1071 } |
| 1072 |
| 1073 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1074 _mm_set_epi64x(long long q1, long long q0) |
| 1075 { |
| 1076 return (__m128i){ q0, q1 }; |
| 1077 } |
| 1078 |
| 1079 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1080 _mm_set_epi64(__m64 q1, __m64 q0) |
| 1081 { |
| 1082 return (__m128i){ (long long)q0, (long long)q1 }; |
| 1083 } |
| 1084 |
| 1085 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1086 _mm_set_epi32(int i3, int i2, int i1, int i0) |
| 1087 { |
| 1088 return (__m128i)(__v4si){ i0, i1, i2, i3}; |
| 1089 } |
| 1090 |
| 1091 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1092 _mm_set_epi16(short w7, short w6, short w5, short w4, short w3, short w2, short
w1, short w0) |
| 1093 { |
| 1094 return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 }; |
| 1095 } |
| 1096 |
| 1097 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1098 _mm_set_epi8(char b15, char b14, char b13, char b12, char b11, char b10, char b9
, char b8, char b7, char b6, char b5, char b4, char b3, char b2, char b1, char b
0) |
| 1099 { |
| 1100 return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b
12, b13, b14, b15 }; |
| 1101 } |
| 1102 |
| 1103 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1104 _mm_set1_epi64x(long long q) |
| 1105 { |
| 1106 return (__m128i){ q, q }; |
| 1107 } |
| 1108 |
| 1109 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1110 _mm_set1_epi64(__m64 q) |
| 1111 { |
| 1112 return (__m128i){ (long long)q, (long long)q }; |
| 1113 } |
| 1114 |
| 1115 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1116 _mm_set1_epi32(int i) |
| 1117 { |
| 1118 return (__m128i)(__v4si){ i, i, i, i }; |
| 1119 } |
| 1120 |
| 1121 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1122 _mm_set1_epi16(short w) |
| 1123 { |
| 1124 return (__m128i)(__v8hi){ w, w, w, w, w, w, w, w }; |
| 1125 } |
| 1126 |
| 1127 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1128 _mm_set1_epi8(char b) |
| 1129 { |
| 1130 return (__m128i)(__v16qi){ b, b, b, b, b, b, b, b, b, b, b, b, b, b, b, b }; |
| 1131 } |
| 1132 |
| 1133 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1134 _mm_setr_epi64(__m64 q0, __m64 q1) |
| 1135 { |
| 1136 return (__m128i){ (long long)q0, (long long)q1 }; |
| 1137 } |
| 1138 |
| 1139 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1140 _mm_setr_epi32(int i0, int i1, int i2, int i3) |
| 1141 { |
| 1142 return (__m128i)(__v4si){ i0, i1, i2, i3}; |
| 1143 } |
| 1144 |
| 1145 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1146 _mm_setr_epi16(short w0, short w1, short w2, short w3, short w4, short w5, short
w6, short w7) |
| 1147 { |
| 1148 return (__m128i)(__v8hi){ w0, w1, w2, w3, w4, w5, w6, w7 }; |
| 1149 } |
| 1150 |
| 1151 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1152 _mm_setr_epi8(char b0, char b1, char b2, char b3, char b4, char b5, char b6, cha
r b7, char b8, char b9, char b10, char b11, char b12, char b13, char b14, char b
15) |
| 1153 { |
| 1154 return (__m128i)(__v16qi){ b0, b1, b2, b3, b4, b5, b6, b7, b8, b9, b10, b11, b
12, b13, b14, b15 }; |
| 1155 } |
| 1156 |
| 1157 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1158 _mm_setzero_si128(void) |
| 1159 { |
| 1160 return (__m128i){ 0LL, 0LL }; |
| 1161 } |
| 1162 |
| 1163 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1164 _mm_store_si128(__m128i *p, __m128i b) |
| 1165 { |
| 1166 *p = b; |
| 1167 } |
| 1168 |
| 1169 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1170 _mm_storeu_si128(__m128i *p, __m128i b) |
| 1171 { |
| 1172 __builtin_ia32_storedqu((char *)p, (__v16qi)b); |
| 1173 } |
| 1174 |
| 1175 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1176 _mm_maskmoveu_si128(__m128i d, __m128i n, char *p) |
| 1177 { |
| 1178 __builtin_ia32_maskmovdqu((__v16qi)d, (__v16qi)n, p); |
| 1179 } |
| 1180 |
| 1181 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1182 _mm_storel_epi64(__m128i *p, __m128i a) |
| 1183 { |
| 1184 __builtin_ia32_storelv4si((__v2si *)p, a); |
| 1185 } |
| 1186 |
| 1187 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1188 _mm_stream_pd(double *p, __m128d a) |
| 1189 { |
| 1190 __builtin_ia32_movntpd(p, a); |
| 1191 } |
| 1192 |
| 1193 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1194 _mm_stream_si128(__m128i *p, __m128i a) |
| 1195 { |
| 1196 __builtin_ia32_movntdq(p, a); |
| 1197 } |
| 1198 |
| 1199 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1200 _mm_stream_si32(int *p, int a) |
| 1201 { |
| 1202 __builtin_ia32_movnti(p, a); |
| 1203 } |
| 1204 |
| 1205 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1206 _mm_clflush(void const *p) |
| 1207 { |
| 1208 __builtin_ia32_clflush(p); |
| 1209 } |
| 1210 |
| 1211 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1212 _mm_lfence(void) |
| 1213 { |
| 1214 __builtin_ia32_lfence(); |
| 1215 } |
| 1216 |
| 1217 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1218 _mm_mfence(void) |
| 1219 { |
| 1220 __builtin_ia32_mfence(); |
| 1221 } |
| 1222 |
| 1223 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1224 _mm_packs_epi16(__m128i a, __m128i b) |
| 1225 { |
| 1226 return (__m128i)__builtin_ia32_packsswb128((__v8hi)a, (__v8hi)b); |
| 1227 } |
| 1228 |
| 1229 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1230 _mm_packs_epi32(__m128i a, __m128i b) |
| 1231 { |
| 1232 return (__m128i)__builtin_ia32_packssdw128((__v4si)a, (__v4si)b); |
| 1233 } |
| 1234 |
| 1235 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1236 _mm_packus_epi16(__m128i a, __m128i b) |
| 1237 { |
| 1238 return (__m128i)__builtin_ia32_packuswb128((__v8hi)a, (__v8hi)b); |
| 1239 } |
| 1240 |
| 1241 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 1242 _mm_extract_epi16(__m128i a, int imm) |
| 1243 { |
| 1244 __v8hi b = (__v8hi)a; |
| 1245 return (unsigned short)b[imm]; |
| 1246 } |
| 1247 |
| 1248 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1249 _mm_insert_epi16(__m128i a, int b, int imm) |
| 1250 { |
| 1251 __v8hi c = (__v8hi)a; |
| 1252 c[imm & 7] = b; |
| 1253 return (__m128i)c; |
| 1254 } |
| 1255 |
| 1256 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 1257 _mm_movemask_epi8(__m128i a) |
| 1258 { |
| 1259 return __builtin_ia32_pmovmskb128((__v16qi)a); |
| 1260 } |
| 1261 |
| 1262 #define _mm_shuffle_epi32(a, imm) \ |
| 1263 ((__m128i)__builtin_shufflevector((__v4si)(a), (__v4si) _mm_set1_epi32(0), \ |
| 1264 (imm) & 0x3, ((imm) & 0xc) >> 2, \ |
| 1265 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6)) |
| 1266 |
| 1267 |
| 1268 #define _mm_shufflelo_epi16(a, imm) \ |
| 1269 ((__m128i)__builtin_shufflevector((__v8hi)(a), (__v8hi) _mm_set1_epi16(0), \ |
| 1270 (imm) & 0x3, ((imm) & 0xc) >> 2, \ |
| 1271 ((imm) & 0x30) >> 4, ((imm) & 0xc0) >> 6, \ |
| 1272 4, 5, 6, 7)) |
| 1273 #define _mm_shufflehi_epi16(a, imm) \ |
| 1274 ((__m128i)__builtin_shufflevector((__v8hi)(a), (__v8hi) _mm_set1_epi16(0), 0,
1, 2, 3, \ |
| 1275 4 + (((imm) & 0x03) >> 0), \ |
| 1276 4 + (((imm) & 0x0c) >> 2), \ |
| 1277 4 + (((imm) & 0x30) >> 4), \ |
| 1278 4 + (((imm) & 0xc0) >> 6))) |
| 1279 |
| 1280 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1281 _mm_unpackhi_epi8(__m128i a, __m128i b) |
| 1282 { |
| 1283 return (__m128i)__builtin_shufflevector((__v16qi)a, (__v16qi)b, 8, 16+8, 9, 16
+9, 10, 16+10, 11, 16+11, 12, 16+12, 13, 16+13, 14, 16+14, 15, 16+15); |
| 1284 } |
| 1285 |
| 1286 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1287 _mm_unpackhi_epi16(__m128i a, __m128i b) |
| 1288 { |
| 1289 return (__m128i)__builtin_shufflevector((__v8hi)a, (__v8hi)b, 4, 8+4, 5, 8+5,
6, 8+6, 7, 8+7); |
| 1290 } |
| 1291 |
| 1292 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1293 _mm_unpackhi_epi32(__m128i a, __m128i b) |
| 1294 { |
| 1295 return (__m128i)__builtin_shufflevector((__v4si)a, (__v4si)b, 2, 4+2, 3, 4+3); |
| 1296 } |
| 1297 |
| 1298 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1299 _mm_unpackhi_epi64(__m128i a, __m128i b) |
| 1300 { |
| 1301 return (__m128i)__builtin_shufflevector(a, b, 1, 2+1); |
| 1302 } |
| 1303 |
| 1304 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1305 _mm_unpacklo_epi8(__m128i a, __m128i b) |
| 1306 { |
| 1307 return (__m128i)__builtin_shufflevector((__v16qi)a, (__v16qi)b, 0, 16+0, 1, 16
+1, 2, 16+2, 3, 16+3, 4, 16+4, 5, 16+5, 6, 16+6, 7, 16+7); |
| 1308 } |
| 1309 |
| 1310 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1311 _mm_unpacklo_epi16(__m128i a, __m128i b) |
| 1312 { |
| 1313 return (__m128i)__builtin_shufflevector((__v8hi)a, (__v8hi)b, 0, 8+0, 1, 8+1,
2, 8+2, 3, 8+3); |
| 1314 } |
| 1315 |
| 1316 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1317 _mm_unpacklo_epi32(__m128i a, __m128i b) |
| 1318 { |
| 1319 return (__m128i)__builtin_shufflevector((__v4si)a, (__v4si)b, 0, 4+0, 1, 4+1); |
| 1320 } |
| 1321 |
| 1322 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1323 _mm_unpacklo_epi64(__m128i a, __m128i b) |
| 1324 { |
| 1325 return (__m128i)__builtin_shufflevector(a, b, 0, 2+0); |
| 1326 } |
| 1327 |
| 1328 static __inline__ __m64 __attribute__((__always_inline__, __nodebug__)) |
| 1329 _mm_movepi64_pi64(__m128i a) |
| 1330 { |
| 1331 return (__m64)a[0]; |
| 1332 } |
| 1333 |
| 1334 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1335 _mm_movpi64_pi64(__m64 a) |
| 1336 { |
| 1337 return (__m128i){ (long long)a, 0 }; |
| 1338 } |
| 1339 |
| 1340 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1341 _mm_move_epi64(__m128i a) |
| 1342 { |
| 1343 return __builtin_shufflevector(a, (__m128i){ 0 }, 0, 2); |
| 1344 } |
| 1345 |
| 1346 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 1347 _mm_unpackhi_pd(__m128d a, __m128d b) |
| 1348 { |
| 1349 return __builtin_shufflevector(a, b, 1, 2+1); |
| 1350 } |
| 1351 |
| 1352 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 1353 _mm_unpacklo_pd(__m128d a, __m128d b) |
| 1354 { |
| 1355 return __builtin_shufflevector(a, b, 0, 2+0); |
| 1356 } |
| 1357 |
| 1358 static __inline__ int __attribute__((__always_inline__, __nodebug__)) |
| 1359 _mm_movemask_pd(__m128d a) |
| 1360 { |
| 1361 return __builtin_ia32_movmskpd(a); |
| 1362 } |
| 1363 |
| 1364 #define _mm_shuffle_pd(a, b, i) \ |
| 1365 (__builtin_shufflevector((__m128d)(a), (__m128d)(b), (i) & 1, \ |
| 1366 (((i) & 2) >> 1) + 2)) |
| 1367 |
| 1368 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 1369 _mm_castpd_ps(__m128d in) |
| 1370 { |
| 1371 return (__m128)in; |
| 1372 } |
| 1373 |
| 1374 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1375 _mm_castpd_si128(__m128d in) |
| 1376 { |
| 1377 return (__m128i)in; |
| 1378 } |
| 1379 |
| 1380 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 1381 _mm_castps_pd(__m128 in) |
| 1382 { |
| 1383 return (__m128d)in; |
| 1384 } |
| 1385 |
| 1386 static __inline__ __m128i __attribute__((__always_inline__, __nodebug__)) |
| 1387 _mm_castps_si128(__m128 in) |
| 1388 { |
| 1389 return (__m128i)in; |
| 1390 } |
| 1391 |
| 1392 static __inline__ __m128 __attribute__((__always_inline__, __nodebug__)) |
| 1393 _mm_castsi128_ps(__m128i in) |
| 1394 { |
| 1395 return (__m128)in; |
| 1396 } |
| 1397 |
| 1398 static __inline__ __m128d __attribute__((__always_inline__, __nodebug__)) |
| 1399 _mm_castsi128_pd(__m128i in) |
| 1400 { |
| 1401 return (__m128d)in; |
| 1402 } |
| 1403 |
| 1404 static __inline__ void __attribute__((__always_inline__, __nodebug__)) |
| 1405 _mm_pause(void) |
| 1406 { |
| 1407 __asm__ volatile ("pause"); |
| 1408 } |
| 1409 |
| 1410 #define _MM_SHUFFLE2(x, y) (((x) << 1) | (y)) |
| 1411 |
| 1412 #endif /* __SSE2__ */ |
| 1413 |
| 1414 #endif /* __EMMINTRIN_H */ |
OLD | NEW |