| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 7/2025. All rights reserved. |
| 3 |
* // |
| 4 |
* // Redistribution and use in source and binary forms, with or without modification, |
| 5 |
* // are permitted provided that the following conditions are met: |
| 6 |
* // |
| 7 |
* // 1. Redistributions of source code must retain the above copyright notice, this |
| 8 |
* // list of conditions and the following disclaimer. |
| 9 |
* // |
| 10 |
* // 2. Redistributions in binary form must reproduce the above copyright notice, |
| 11 |
* // this list of conditions and the following disclaimer in the documentation |
| 12 |
* // and/or other materials provided with the distribution. |
| 13 |
* // |
| 14 |
* // 3. Neither the name of the copyright holder nor the names of its |
| 15 |
* // contributors may be used to endorse or promote products derived from |
| 16 |
* // this software without specific prior written permission. |
| 17 |
* // |
| 18 |
* // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" |
| 19 |
* // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE |
| 20 |
* // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE |
| 21 |
* // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE |
| 22 |
* // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL |
| 23 |
* // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR |
| 24 |
* // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER |
| 25 |
* // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, |
| 26 |
* // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE |
| 27 |
* // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. |
| 28 |
*/ |
| 29 |
use crate::common::dd_fmla; |
| 30 |
use crate::double_double::DoubleDouble; |
| 31 |
use crate::err::erf::{Erf, erf_accurate, erf_fast}; |
| 32 |
use crate::exponents::{EXP_REDUCE_T0, EXP_REDUCE_T1, ldexp}; |
| 33 |
use crate::rounding::CpuRoundTiesEven; |
| 34 |
use std::hint::black_box; |
| 35 |
|
| 36 |
static ASYMPTOTIC_POLY: [[u64; 13]; 6] = [ |
| 37 |
[ |
| 38 |
0x3fe20dd750429b6d, |
| 39 |
0x3c61a1feb75a48a8, |
| 40 |
0xbfd20dd750429b6c, |
| 41 |
0x3fdb14c2f863e403, |
| 42 |
0xbff0ecf9db3af35d, |
| 43 |
0x400d9eb53ca6eeed, |
| 44 |
0xc030a945830d95c8, |
| 45 |
0x4056e8a963e2f1f5, |
| 46 |
0xc0829b7ccc8f396f, |
| 47 |
0x40b15e716e83c27e, |
| 48 |
0xc0e1cfdcfbcaf22a, |
| 49 |
0x4111986cc7a7e8fe, |
| 50 |
0xc1371f7540590a91, |
| 51 |
], |
| 52 |
[ |
| 53 |
0x3fe20dd750429ae7, |
| 54 |
0x3c863da89e801fd4, |
| 55 |
0xbfd20dd750400795, |
| 56 |
0x3fdb14c2f57c490c, |
| 57 |
0xbff0ecf95c8c9014, |
| 58 |
0x400d9e981f2321ef, |
| 59 |
0xc030a81482de1506, |
| 60 |
0x4056d662420a604b, |
| 61 |
0xc08233c96fff7772, |
| 62 |
0x40af5d62018d3e37, |
| 63 |
0xc0d9ae55e9554450, |
| 64 |
0x410052901e10d139, |
| 65 |
0xc1166465df1385f0, |
| 66 |
], |
| 67 |
[ |
| 68 |
0x3fe20dd75041e3fc, |
| 69 |
0xbc7c9b491c4920fc, |
| 70 |
0xbfd20dd74e5f1526, |
| 71 |
0x3fdb14c1d35a40e0, |
| 72 |
0xbff0ecdecd30e86b, |
| 73 |
0x400d9b4e7f725263, |
| 74 |
0xc030958b5ca8fb39, |
| 75 |
0x40563e3179bf609c, |
| 76 |
0xc0806bbd1cd2d0fd, |
| 77 |
0x40a7b66eb6d1d2f2, |
| 78 |
0xc0cce5a4b1afab75, |
| 79 |
0x40e8b5c6ae6f773c, |
| 80 |
0xc0f5475860326f86, |
| 81 |
], |
| 82 |
[ |
| 83 |
0x3fe20dd75025cfe9, |
| 84 |
0x3c55a92eef32fb20, |
| 85 |
0xbfd20dd71eb9d4e7, |
| 86 |
0x3fdb14af4c25db28, |
| 87 |
0xbff0ebc78a22b3d8, |
| 88 |
0x400d85287a0b3399, |
| 89 |
0xc03045f751e5ca1d, |
| 90 |
0x4054a0d87ddea589, |
| 91 |
0xc07ac6a0981d1eee, |
| 92 |
0x409f44822f567956, |
| 93 |
0xc0bcba372de71349, |
| 94 |
0x40d1a4a19f550ca4, |
| 95 |
0xc0d52a580455ed79, |
| 96 |
], |
| 97 |
[ |
| 98 |
0x3fe20dd74eb31d84, |
| 99 |
0xbc439c4054b7c090, |
| 100 |
0xbfd20dd561af98c4, |
| 101 |
0x3fdb1435165d9df1, |
| 102 |
0xbff0e6b60308e940, |
| 103 |
0x400d3ce30c140882, |
| 104 |
0xc02f2083e404c299, |
| 105 |
0x40520f113d89b42a, |
| 106 |
0xc0741433ebd89f19, |
| 107 |
0x4092f35b6a3154f6, |
| 108 |
0xc0ab020a4313cf3b, |
| 109 |
0x40b90f07e92da7ee, |
| 110 |
0xc0b6565e1d7665c3, |
| 111 |
], |
| 112 |
[ |
| 113 |
0x3fe20dd744b3517b, |
| 114 |
0xbc6f77ab25e01ab4, |
| 115 |
0xbfd20dcc62ec4024, |
| 116 |
0x3fdb125bfa4f66c1, |
| 117 |
0xbff0d80e65381970, |
| 118 |
0x400ca11fbcfa65b2, |
| 119 |
0xc02cd9eaffb88315, |
| 120 |
0x404e010db42e0da7, |
| 121 |
0xc06c5c85250ef6a3, |
| 122 |
0x4085e118d9c1eeaf, |
| 123 |
0xc098d74be13d3d30, |
| 124 |
0x40a211b1b2b5ac83, |
| 125 |
0xc09900be759fc663, |
| 126 |
], |
| 127 |
]; |
| 128 |
|
| 129 |
static ASYMPTOTIC_POLY_ACCURATE: [[u64; 30]; 10] = [ |
| 130 |
[ |
| 131 |
0x3fe20dd750429b6d, |
| 132 |
0x3c61ae3a912b08f0, |
| 133 |
0xbfd20dd750429b6d, |
| 134 |
0xbc51ae34c0606d68, |
| 135 |
0x3fdb14c2f863e924, |
| 136 |
0xbc796c0f4c848fc8, |
| 137 |
0xbff0ecf9db3e71b6, |
| 138 |
0x3c645d756bd288b0, |
| 139 |
0x400d9eb53fad4672, |
| 140 |
0xbcac61629de9adf2, |
| 141 |
0xc030a945f3d147ea, |
| 142 |
0x3cb8fec5ad7ece20, |
| 143 |
0x4056e8c02f27ca6d, |
| 144 |
0xc0829d1c21c363e0, |
| 145 |
0x40b17349b70be627, |
| 146 |
0xc0e28a6bb4686182, |
| 147 |
0x411602d1662523ca, |
| 148 |
0xc14ccae7625c4111, |
| 149 |
0x4184237d064f6e0d, |
| 150 |
0xc1bb1e5466ca3a2f, |
| 151 |
0x41e90ae06a0f6cc1, |
| 152 |
0, |
| 153 |
0, |
| 154 |
0, |
| 155 |
0, |
| 156 |
0, |
| 157 |
0, |
| 158 |
0, |
| 159 |
0, |
| 160 |
0, |
| 161 |
], |
| 162 |
[ |
| 163 |
0x3fe20dd750429b6d, |
| 164 |
0x3c61adaa62435c10, |
| 165 |
0xbfd20dd750429b6d, |
| 166 |
0xbc441516126827c8, |
| 167 |
0x3fdb14c2f863e90b, |
| 168 |
0x3c7a535780ba5ed4, |
| 169 |
0xbff0ecf9db3e65d6, |
| 170 |
0xbc9089edde27ad07, |
| 171 |
0x400d9eb53fa52f20, |
| 172 |
0xbcabc9737e9464ac, |
| 173 |
0xc030a945f2cd7621, |
| 174 |
0xbcc589f28b700332, |
| 175 |
0x4056e8bffd7e194e, |
| 176 |
0xc0829d18716876e2, |
| 177 |
0x40b17312abe18250, |
| 178 |
0xc0e287e73592805c, |
| 179 |
0x4115ebf7394a39c1, |
| 180 |
0xc14c2f14d46d0cf9, |
| 181 |
0x4182af3d256f955e, |
| 182 |
0xc1b7041659ebd7aa, |
| 183 |
0x41e6039c232e2f71, |
| 184 |
0xc2070ca15c5a07cb, |
| 185 |
0, |
| 186 |
0, |
| 187 |
0, |
| 188 |
0, |
| 189 |
0, |
| 190 |
0, |
| 191 |
0, |
| 192 |
0, |
| 193 |
], |
| 194 |
[ |
| 195 |
0x3fe20dd750429b6d, |
| 196 |
0x3c5d3c35b5d37410, |
| 197 |
0xbfd20dd750429b56, |
| 198 |
0xbc7c028415f6f81b, |
| 199 |
0x3fdb14c2f863c1cf, |
| 200 |
0x3c51bb0de6470dbc, |
| 201 |
0xbff0ecf9db33c363, |
| 202 |
0x3c80f8068459eb16, |
| 203 |
0x400d9eb53b9ce57b, |
| 204 |
0x3ca20cce33e7d84a, |
| 205 |
0xc030a945aa2ec4fa, |
| 206 |
0xbcdf6e0fcd7c6030, |
| 207 |
0x4056e8b824d2bfaa, |
| 208 |
0xc0829cc372a6d0b0, |
| 209 |
0x40b1703a99ddd429, |
| 210 |
0xc0e2749f9a267cc6, |
| 211 |
0x4115856a17271849, |
| 212 |
0xc14a8bcb4ba9753f, |
| 213 |
0x418035dcce882940, |
| 214 |
0xc1b1e5d8c5e6e043, |
| 215 |
0x41dfe3b4f365386e, |
| 216 |
0xc20398fdef2b98fe, |
| 217 |
0x42184234d4f4ea12, |
| 218 |
0, |
| 219 |
0, |
| 220 |
0, |
| 221 |
0, |
| 222 |
0, |
| 223 |
0, |
| 224 |
0, |
| 225 |
], |
| 226 |
[ |
| 227 |
0x3fe20dd750429b6a, |
| 228 |
0x3c8ae622b765e9fd, |
| 229 |
0xbfd20dd750428f0e, |
| 230 |
0x3c703c6c67d69513, |
| 231 |
0x3fdb14c2f8563e8e, |
| 232 |
0x3c6766a6bd7aa89c, |
| 233 |
0xbff0ecf9d8dedd48, |
| 234 |
0x3c90af52e90336e3, |
| 235 |
0x400d9eb4aad086fe, |
| 236 |
0x3ca640d371d54a19, |
| 237 |
0xc030a93f1d01cfe0, |
| 238 |
0xbcc68dbd8d9c522c, |
| 239 |
0x4056e842e9fd5898, |
| 240 |
0xc08299886ef1fb80, |
| 241 |
0x40b15e0f0162c9a0, |
| 242 |
0xc0e222dbc6b04cd8, |
| 243 |
0x411460268db1ebdf, |
| 244 |
0xc1474f53ce065fb3, |
| 245 |
0x417961ca8553f870, |
| 246 |
0xc1a8788395d13798, |
| 247 |
0x41d35e37b25d0e81, |
| 248 |
0xc1f707b7457c8f5e, |
| 249 |
0x4211ff852df1c023, |
| 250 |
0xc21b75d0ec56e2cd, |
| 251 |
0, |
| 252 |
0, |
| 253 |
0, |
| 254 |
0, |
| 255 |
0, |
| 256 |
0, |
| 257 |
], |
| 258 |
[ |
| 259 |
0x3fe20dd750429a8f, |
| 260 |
0xbc766d8dda59bcea, |
| 261 |
0xbfd20dd7503fdbab, |
| 262 |
0x3c6707bdffc2b3fe, |
| 263 |
0x3fdb14c2f6526025, |
| 264 |
0xbc27fa4bb9541140, |
| 265 |
0xbff0ecf99c417d45, |
| 266 |
0xbc9748645ef7af94, |
| 267 |
0x400d9eaa9c712a7d, |
| 268 |
0x3ca79e478994ebb4, |
| 269 |
0xc030a8ef11fbf141, |
| 270 |
0x3cbb5c72d69f8954, |
| 271 |
0x4056e4653e0455b1, |
| 272 |
0xc08286909448e6cf, |
| 273 |
0x40b113424ce76821, |
| 274 |
0xc0e1346d859e76de, |
| 275 |
0x4111f9f6cf2293bf, |
| 276 |
0xc14258e6e3b337db, |
| 277 |
0x41714029ecd465fb, |
| 278 |
0xc19c530df5337a6f, |
| 279 |
0x41c34bc4bbccd336, |
| 280 |
0xc1e4a37c52641688, |
| 281 |
0x420019707cec2974, |
| 282 |
0xc21031fe736ea169, |
| 283 |
0x420f6b3003de3ddf, |
| 284 |
0, |
| 285 |
0, |
| 286 |
0, |
| 287 |
0, |
| 288 |
0, |
| 289 |
], |
| 290 |
[ |
| 291 |
0x3fe20dd75042756b, |
| 292 |
0x3c84ad9178b56910, |
| 293 |
0xbfd20dd74feda9e8, |
| 294 |
0xbc78141c70bbc8d6, |
| 295 |
0x3fdb14c2cb128467, |
| 296 |
0xbc709aebaa106821, |
| 297 |
0xbff0ecf603921a0b, |
| 298 |
0x3c97d3cb5bceaf0b, |
| 299 |
0x400d9e3e1751ca59, |
| 300 |
0x3c76622ae5642670, |
| 301 |
0xc030a686af57f547, |
| 302 |
0x3cc083b320aff6b6, |
| 303 |
0x4056cf0b6c027326, |
| 304 |
0xc0823afcb69443d3, |
| 305 |
0x40b03ab450d9f1b9, |
| 306 |
0xc0de74cdb76bcab4, |
| 307 |
0x410c671b60e607f1, |
| 308 |
0xc138f1376d324ce4, |
| 309 |
0x4163b64276234676, |
| 310 |
0xc18aff0ce13c5a8e, |
| 311 |
0x41aef20247251e87, |
| 312 |
0xc1cc9f5662f721f6, |
| 313 |
0x41e4687858e185e1, |
| 314 |
0xc1f4fa507be073c2, |
| 315 |
0x41fb99ac35ee4acc, |
| 316 |
0xc1f16cb585ee3fa9, |
| 317 |
0, |
| 318 |
0, |
| 319 |
0, |
| 320 |
0, |
| 321 |
], |
| 322 |
[ |
| 323 |
0x3fe20dd7503e730d, |
| 324 |
0x3c84e524a098a467, |
| 325 |
0xbfd20dd7498fa6b2, |
| 326 |
0x3c260a4e27751c80, |
| 327 |
0x3fdb14c061bd2a0c, |
| 328 |
0x3c695a8f847d2fc2, |
| 329 |
0xbff0ecd0f11b8c7d, |
| 330 |
0xbc94126deea76061, |
| 331 |
0x400d9b1344463548, |
| 332 |
0x3cafe09a4eca9b0e, |
| 333 |
0xc030996ea52a87ed, |
| 334 |
0xbca924f920db26c0, |
| 335 |
0x40567a2264b556b0, |
| 336 |
0xc0815dfc2c86b6b5, |
| 337 |
0x40accc291b62efe4, |
| 338 |
0xc0d81375a78e746a, |
| 339 |
0x41033a6f15546329, |
| 340 |
0xc12c1e9dc1216010, |
| 341 |
0x4152397ea3d43fda, |
| 342 |
0xc174661e5b2ea512, |
| 343 |
0x4193412367ca5d45, |
| 344 |
0xc1ade56b9d7f37c4, |
| 345 |
0x41c2851d9722146d, |
| 346 |
0xc1d19027baf0c3fe, |
| 347 |
0x41d7e7b8b6ab58ac, |
| 348 |
0xc1d4c446d56aaf22, |
| 349 |
0x41c1492190400505, |
| 350 |
0, |
| 351 |
0, |
| 352 |
0, |
| 353 |
], |
| 354 |
[ |
| 355 |
0x3fe20dd74ff10852, |
| 356 |
0x3c8a32f26deff875, |
| 357 |
0xbfd20dd6f06c491c, |
| 358 |
0x3c770c16e1793358, |
| 359 |
0x3fdb14a7d5e7fd4a, |
| 360 |
0x3c7479998b54db5b, |
| 361 |
0xbff0ebbdb3889c5f, |
| 362 |
0xbc759b853e11369c, |
| 363 |
0x400d89dd249d7ef8, |
| 364 |
0xbc84b5edf0c8c314, |
| 365 |
0xc0306526fb386114, |
| 366 |
0xbc840d04eed7c7e0, |
| 367 |
0x40557ff657e429ce, |
| 368 |
0xc07ef63e90d38630, |
| 369 |
0x40a6d4f34c4ea3da, |
| 370 |
0xc0d04542b9e36a54, |
| 371 |
0x40f577bf19097738, |
| 372 |
0xc119702fe47c736d, |
| 373 |
0x413a7ae12b54fdc6, |
| 374 |
0xc157ca3f0f7c4fa9, |
| 375 |
0x417225d983963cbf, |
| 376 |
0xc1871a6eac612f9e, |
| 377 |
0x4198086324225e1e, |
| 378 |
0xc1a3de68670a7716, |
| 379 |
0x41a91674de4dcbe9, |
| 380 |
0xc1a6b44cc15b76c2, |
| 381 |
0x419a36dae0f30d80, |
| 382 |
0xc17cffc1747ea3dc, |
| 383 |
0, |
| 384 |
0, |
| 385 |
], |
| 386 |
[ |
| 387 |
0x3fe20dd74ba8f300, |
| 388 |
0xbc59dd256871d210, |
| 389 |
0xbfd20dd3593675bc, |
| 390 |
0x3c7ec0e7ffa91ad9, |
| 391 |
0x3fdb13eef86a077a, |
| 392 |
0xbc74fb5d78d411b8, |
| 393 |
0xbff0e5cf52a11f3a, |
| 394 |
0xbc851f36c779dc8c, |
| 395 |
0x400d4417a08b39d5, |
| 396 |
0x3c91be9fb5956638, |
| 397 |
0xc02f91b9f6ce80c3, |
| 398 |
0xbccc9c99dd42829c, |
| 399 |
0x405356439f45bb43, |
| 400 |
0xc078c0ca12819b48, |
| 401 |
0x409efcad2ecd6671, |
| 402 |
0xc0c21b0af6fc1039, |
| 403 |
0x40e327d215ee30c9, |
| 404 |
0xc101fabda96167b0, |
| 405 |
0x411d82e4373b315d, |
| 406 |
0xc134ed9e2ff591e9, |
| 407 |
0x41495c85dcd8eab5, |
| 408 |
0xc159f016f0a3d62a, |
| 409 |
0x41660e89d918b96f, |
| 410 |
0xc16e97be202cba64, |
| 411 |
0x4170d8a081619793, |
| 412 |
0xc16c5422b4fcfc65, |
| 413 |
0x4161131a9dc6aed1, |
| 414 |
0xc14a457d9dced257, |
| 415 |
0x4123605e980e8b86, |
| 416 |
0, |
| 417 |
], |
| 418 |
[ |
| 419 |
0x3fe20dd7319d4d25, |
| 420 |
0x3c82b02992c3b7ab, |
| 421 |
0xbfd20dc29c13ab1b, |
| 422 |
0xbc7d78d79b4ad767, |
| 423 |
0x3fdb115a57b5ab13, |
| 424 |
0xbc6aa8c45be0aa2e, |
| 425 |
0xbff0d58ec437efd7, |
| 426 |
0xbc5994f00a15e850, |
| 427 |
0x400cb1742e229f23, |
| 428 |
0xbca8000471d54399, |
| 429 |
0xc02d99a5edf7b946, |
| 430 |
0xbcbaf76ed7e35cde, |
| 431 |
0x4050a8b71058eb28, |
| 432 |
0xc072d88289da5bfc, |
| 433 |
0x40943ddf24168edb, |
| 434 |
0xc0b3e9dfc38b6d1a, |
| 435 |
0x40d18d4df97ab3df, |
| 436 |
0xc0eb550fc62dcab5, |
| 437 |
0x41029cb71f116ed1, |
| 438 |
0xc115fc9cc4e854e3, |
| 439 |
0x41265915fd0567b1, |
| 440 |
0xc1335eb5fca0e46d, |
| 441 |
0x413c5261ecc0d789, |
| 442 |
0xc14138932dc4eafc, |
| 443 |
0x414117d4eb18facd, |
| 444 |
0xc13af96163e35eca, |
| 445 |
0x4130454a3a63c766, |
| 446 |
0xc11c2ebc1d39b44a, |
| 447 |
0x40ff3327698e0e6b, |
| 448 |
0xc0d094febc3dff35, |
| 449 |
], |
| 450 |
]; |
| 451 |
|
| 452 |
// Approximation for the fast path of exp(z) for z=zh+zl, |
| 453 |
// with |z| < 0.000130273 < 2^-12.88 and |zl| < 2^-42.6 |
| 454 |
// (assuming x^y does not overflow or underflow) |
| 455 |
#[inline] |
| 456 |
fn q_1(z_dd: DoubleDouble) -> DoubleDouble { |
| 457 |
const C: [u64; 5] = [ |
| 458 |
0x3ff0000000000000, |
| 459 |
0x3ff0000000000000, |
| 460 |
0x3fe0000000000000, |
| 461 |
0x3fc5555555995d37, |
| 462 |
0x3fa55555558489dc, |
| 463 |
]; |
| 464 |
let z = z_dd.to_f64(); |
| 465 |
let mut q = dd_fmla(f64::from_bits(C[4]), z_dd.hi, f64::from_bits(C[3])); |
| 466 |
|
| 467 |
q = dd_fmla(q, z, f64::from_bits(C[2])); |
| 468 |
|
| 469 |
let mut v = DoubleDouble::from_exact_add(f64::from_bits(C[1]), q * z); |
| 470 |
v = DoubleDouble::quick_mult(z_dd, v); |
| 471 |
DoubleDouble::f64_add(f64::from_bits(C[0]), v) |
| 472 |
} |
| 473 |
|
| 474 |
#[inline] |
| 475 |
fn exp_1(x: DoubleDouble) -> DoubleDouble { |
| 476 |
const INVLOG2: f64 = f64::from_bits(0x40b71547652b82fe); /* |INVLOG2-2^12/log(2)| < 2^-43.4 */ |
| 477 |
let k = (x.hi * INVLOG2).cpu_round_ties_even(); |
| 478 |
|
| 479 |
const LOG2_DD: DoubleDouble = DoubleDouble::new( |
| 480 |
f64::from_bits(0x3bbabc9e3b39803f), |
| 481 |
f64::from_bits(0x3f262e42fefa39ef), |
| 482 |
); |
| 483 |
let k_dd = DoubleDouble::quick_f64_mult(k, LOG2_DD); |
| 484 |
let mut y_dd = DoubleDouble::from_exact_add(x.hi - k_dd.hi, x.lo); |
| 485 |
y_dd.lo -= k_dd.lo; |
| 486 |
|
| 487 |
let ki: i64 = k as i64; /* Note: k is an integer, this is just a conversion. */ |
| 488 |
let mi = (ki >> 12).wrapping_add(0x3ff); |
| 489 |
let i2: i64 = (ki >> 6) & 0x3f; |
| 490 |
let i1: i64 = ki & 0x3f; |
| 491 |
let t1 = DoubleDouble::new( |
| 492 |
f64::from_bits(EXP_REDUCE_T0[i2 as usize].0), |
| 493 |
f64::from_bits(EXP_REDUCE_T0[i2 as usize].1), |
| 494 |
); |
| 495 |
let t2 = DoubleDouble::new( |
| 496 |
f64::from_bits(EXP_REDUCE_T1[i1 as usize].0), |
| 497 |
f64::from_bits(EXP_REDUCE_T1[i1 as usize].1), |
| 498 |
); |
| 499 |
let mut v = DoubleDouble::quick_mult(t2, t1); |
| 500 |
let q = q_1(y_dd); |
| 501 |
v = DoubleDouble::quick_mult(v, q); |
| 502 |
|
| 503 |
let scale = f64::from_bits((mi as u64) << 52); |
| 504 |
v.hi *= scale; |
| 505 |
v.lo *= scale; |
| 506 |
v |
| 507 |
} |
| 508 |
|
| 509 |
struct Exp { |
| 510 |
e: i32, |
| 511 |
result: DoubleDouble, |
| 512 |
} |
| 513 |
|
| 514 |
fn exp_accurate(x_dd: DoubleDouble) -> Exp { |
| 515 |
static E2: [u64; 28] = [ |
| 516 |
0x3ff0000000000000, |
| 517 |
0xb960000000000000, |
| 518 |
0x3ff0000000000000, |
| 519 |
0xb9be200000000000, |
| 520 |
0x3fe0000000000000, |
| 521 |
0x3a03c00000000000, |
| 522 |
0x3fc5555555555555, |
| 523 |
0x3c655555555c78d9, |
| 524 |
0x3fa5555555555555, |
| 525 |
0x3c455555545616e2, |
| 526 |
0x3f81111111111111, |
| 527 |
0x3c011110121fc314, |
| 528 |
0x3f56c16c16c16c17, |
| 529 |
0xbbef49e06ee3a56e, |
| 530 |
0x3f2a01a01a01a01a, |
| 531 |
0x3b6b053e1eeab9c0, |
| 532 |
0x3efa01a01a01a01a, |
| 533 |
0x3ec71de3a556c733, |
| 534 |
0x3e927e4fb7789f66, |
| 535 |
0x3e5ae64567f54abe, |
| 536 |
0x3e21eed8eff8958b, |
| 537 |
0x3de6124613837216, |
| 538 |
0x3da93974aaf26a57, |
| 539 |
0x3d6ae7f4fd6d0bd9, |
| 540 |
0x3d2ae7e982620b25, |
| 541 |
0x3ce94e4ca59460d8, |
| 542 |
0x3ca69a2a4b7ef36d, |
| 543 |
0x3c6abfe1602308c9, |
| 544 |
]; |
| 545 |
const LOG2INV: f64 = f64::from_bits(0x3ff71547652b82fe); |
| 546 |
let k: i32 = unsafe { |
| 547 |
(x_dd.hi * LOG2INV) |
| 548 |
.cpu_round_ties_even() |
| 549 |
.to_int_unchecked::<i32>() |
| 550 |
}; |
| 551 |
|
| 552 |
const LOG2_H: f64 = f64::from_bits(0x3fe62e42fefa39ef); |
| 553 |
/* we approximate LOG2Lacc ~ log(2) - LOG2H with 38 bits, so that |
| 554 |
k*LOG2Lacc is exact (k has at most 11 bits) */ |
| 555 |
const LOG2_L: f64 = f64::from_bits(0x3c7abc9e3b398000); |
| 556 |
const LOG2_TINY: f64 = f64::from_bits(0x398f97b57a079a19); |
| 557 |
let yh = dd_fmla(-k as f64, LOG2_H, x_dd.hi); |
| 558 |
/* since |xh+xl| >= 2.92 we have |k| >= 4; |
| 559 |
(|k|-1/2)*log(2) <= |x| <= (|k|+1/2)*log(2) thus |
| 560 |
1-1/(2|k|) <= |x/(k*log(2))| <= 1+1/(2|k|) thus by Sterbenz theorem |
| 561 |
yh is exact too */ |
| 562 |
let mut t = DoubleDouble::from_full_exact_add(-k as f64 * LOG2_L, x_dd.lo); |
| 563 |
let mut y_dd = DoubleDouble::from_exact_add(yh, t.hi); |
| 564 |
y_dd.lo = dd_fmla(-k as f64, LOG2_TINY, y_dd.lo + t.lo); |
| 565 |
/* now yh+yl approximates xh + xl - k*log(2), and we approximate p(yh+yl) |
| 566 |
in h + l */ |
| 567 |
/* Since |xh| <= 742, we assume |xl| <= ulp(742) = 2^-43. Then since |
| 568 |
|k| <= round(742/log(2)) = 1070, |yl| <= 1070*LOG2L + 2^-42 < 2^-42.7. |
| 569 |
Since |yh| <= log(2)/2, the contribution of yl is negligible as long |
| 570 |
as |i*p[i]*yh^(i-1)*yl| < 2^-104, which holds for i >= 16. |
| 571 |
Thus for coefficients of degree 16 or more, we don't take yl into account. |
| 572 |
*/ |
| 573 |
let mut h = f64::from_bits(E2[19 + 8]); // degree 19 |
| 574 |
for a in (16..=18).rev() { |
| 575 |
h = dd_fmla(h, y_dd.hi, f64::from_bits(E2[a + 8])); // degree i |
| 576 |
} |
| 577 |
/* degree 15: h*(yh+yl)+E2[15 + 8] */ |
| 578 |
t = DoubleDouble::from_exact_mult(h, y_dd.hi); |
| 579 |
t.lo = dd_fmla(h, y_dd.lo, t.lo); |
| 580 |
let mut v = DoubleDouble::from_exact_add(f64::from_bits(E2[15 + 8]), t.hi); |
| 581 |
v.lo += t.lo; |
| 582 |
for a in (8..=14).rev() { |
| 583 |
/* degree i: (h+l)*(yh+yl)+E2[i+8] */ |
| 584 |
t = DoubleDouble::quick_mult(v, y_dd); |
| 585 |
v = DoubleDouble::from_exact_add(f64::from_bits(E2[a + 8]), t.hi); |
| 586 |
v.lo += t.lo; |
| 587 |
} |
| 588 |
for a in (0..=7).rev() { |
| 589 |
/* degree i: (h+l)*(yh+yl)+E2[2i]+E2[2i+1] */ |
| 590 |
t = DoubleDouble::quick_mult(v, y_dd); |
| 591 |
v = DoubleDouble::from_exact_add(f64::from_bits(E2[2 * a]), t.hi); |
| 592 |
v.lo += t.lo + f64::from_bits(E2[2 * a + 1]); |
| 593 |
} |
| 594 |
|
| 595 |
Exp { e: k, result: v } |
| 596 |
} |
| 597 |
|
| 598 |
#[cold] |
| 599 |
fn erfc_asympt_accurate(x: f64) -> f64 { |
| 600 |
/* subnormal exceptions */ |
| 601 |
if x == f64::from_bits(0x403a8f7bfbd15495) { |
| 602 |
return dd_fmla( |
| 603 |
f64::from_bits(0x0000000000000001), |
| 604 |
-0.25, |
| 605 |
f64::from_bits(0x000667bd620fd95b), |
| 606 |
); |
| 607 |
} |
| 608 |
let u_dd = DoubleDouble::from_exact_mult(x, x); |
| 609 |
let exp_result = exp_accurate(DoubleDouble::new(-u_dd.lo, -u_dd.hi)); |
| 610 |
|
| 611 |
/* compute 1/x as double-double */ |
| 612 |
let yh = 1.0 / x; |
| 613 |
/* Newton's iteration for 1/x is y -> y + y*(1-x*y) */ |
| 614 |
let yl = yh * dd_fmla(-x, yh, 1.0); |
| 615 |
// yh+yl approximates 1/x |
| 616 |
static THRESHOLD: [u64; 10] = [ |
| 617 |
0x3fb4500000000000, |
| 618 |
0x3fbe000000000000, |
| 619 |
0x3fc3f00000000000, |
| 620 |
0x3fc9500000000000, |
| 621 |
0x3fcf500000000000, |
| 622 |
0x3fd3100000000000, |
| 623 |
0x3fd7100000000000, |
| 624 |
0x3fdbc00000000000, |
| 625 |
0x3fe0b00000000000, |
| 626 |
0x3fe3000000000000, |
| 627 |
]; |
| 628 |
let mut i = 0usize; |
| 629 |
while i < THRESHOLD.len() && yh > f64::from_bits(THRESHOLD[i]) { |
| 630 |
i += 1; |
| 631 |
} |
| 632 |
let p = ASYMPTOTIC_POLY_ACCURATE[i]; |
| 633 |
let mut u_dd = DoubleDouble::from_exact_mult(yh, yh); |
| 634 |
u_dd.lo = dd_fmla(2.0 * yh, yl, u_dd.lo); |
| 635 |
/* the polynomial p has degree 29+2i, and its coefficient of largest |
| 636 |
degree is p[14+6+i] */ |
| 637 |
let mut z_dd = DoubleDouble::new(0., f64::from_bits(p[14 + 6 + i])); |
| 638 |
for a in (13..=27 + 2 * i).rev().step_by(2) { |
| 639 |
/* degree j: (zh+zl)*(uh+ul)+p[(j-1)/2+6]] */ |
| 640 |
let v = DoubleDouble::quick_mult(z_dd, u_dd); |
| 641 |
z_dd = DoubleDouble::from_full_exact_add(f64::from_bits(p[(a - 1) / 2 + 6]), v.hi); |
| 642 |
z_dd.lo += v.lo; |
| 643 |
} |
| 644 |
for a in (1..=11).rev().step_by(2) { |
| 645 |
let v = DoubleDouble::quick_mult(z_dd, u_dd); |
| 646 |
z_dd = DoubleDouble::from_full_exact_add(f64::from_bits(p[a - 1]), v.hi); |
| 647 |
z_dd.lo += v.lo + f64::from_bits(p[a]); |
| 648 |
} |
| 649 |
|
| 650 |
/* multiply by yh+yl */ |
| 651 |
u_dd = DoubleDouble::quick_mult(z_dd, DoubleDouble::new(yl, yh)); |
| 652 |
/* now uh+ul approximates p(1/x), i.e., erfc(x)*exp(x^2) */ |
| 653 |
/* now multiply (uh+ul)*(eh+el), after normalizing uh+ul to reduce the |
| 654 |
number of exceptional cases */ |
| 655 |
u_dd = DoubleDouble::from_exact_add(u_dd.hi, u_dd.lo); |
| 656 |
let v = DoubleDouble::quick_mult(u_dd, exp_result.result); |
| 657 |
/* multiply by 2^e */ |
| 658 |
/* multiply by 2^e */ |
| 659 |
let mut res = ldexp(v.to_f64(), exp_result.e); |
| 660 |
if res < f64::from_bits(0x0010000000000000) { |
| 661 |
/* for erfc(x) in the subnormal range, we have to perform a special |
| 662 |
rounding */ |
| 663 |
let mut corr = v.hi - ldexp(res, -exp_result.e); |
| 664 |
corr += v.lo; |
| 665 |
/* add corr*2^e */ |
| 666 |
res += ldexp(corr, exp_result.e); |
| 667 |
} |
| 668 |
res |
| 669 |
} |
| 670 |
|
| 671 |
#[cold] |
| 672 |
fn erfc_accurate(x: f64) -> f64 { |
| 673 |
if x < 0. { |
| 674 |
let mut v_dd = erf_accurate(-x); |
| 675 |
let t = DoubleDouble::from_exact_add(1.0, v_dd.hi); |
| 676 |
v_dd.hi = t.hi; |
| 677 |
v_dd.lo += t.lo; |
| 678 |
return v_dd.to_f64(); |
| 679 |
} else if x <= f64::from_bits(0x3ffb59ffb450828c) { |
| 680 |
// erfc(x) >= 2^-6 |
| 681 |
let mut v_dd = erf_accurate(x); |
| 682 |
let t = DoubleDouble::from_exact_add(1.0, -v_dd.hi); |
| 683 |
v_dd.hi = t.hi; |
| 684 |
v_dd.lo = t.lo - v_dd.lo; |
| 685 |
return v_dd.to_f64(); |
| 686 |
} |
| 687 |
// now 0x1.b59ffb450828cp+0 < x < 0x1.b39dc41e48bfdp+4 |
| 688 |
erfc_asympt_accurate(x) |
| 689 |
} |
| 690 |
|
| 691 |
/* Fast path for 0x1.713786d9c7c09p+1 < x < 0x1.b39dc41e48bfdp+4, |
| 692 |
using the asymptotic formula erfc(x) = exp(-x^2) * p(1/x)*/ |
| 693 |
fn erfc_asympt_fast(x: f64) -> Erf { |
| 694 |
/* for x >= 0x1.9db1bb14e15cap+4, erfc(x) < 2^-970, and we might encounter |
| 695 |
underflow issues in the computation of l, thus we delegate this case |
| 696 |
to the accurate path */ |
| 697 |
if x >= f64::from_bits(0x4039db1bb14e15ca) { |
| 698 |
return Erf { |
| 699 |
err: 1.0, |
| 700 |
result: DoubleDouble::default(), |
| 701 |
}; |
| 702 |
} |
| 703 |
|
| 704 |
let mut u = DoubleDouble::from_exact_mult(x, x); |
| 705 |
let e_dd = exp_1(DoubleDouble::new(-u.lo, -u.hi)); |
| 706 |
|
| 707 |
/* the assumptions from exp_1 are satisfied: |
| 708 |
* a_mul ensures |ul| <= ulp(uh), thus |ul/uh| <= 2^-52 |
| 709 |
* since |x| < 0x1.9db1bb14e15cap+4 we have |
| 710 |
|ul| < ulp(0x1.9db1bb14e15cap+4^2) = 2^-43 */ |
| 711 |
/* eh+el approximates exp(-x^2) with maximal relative error 2^-74.139 */ |
| 712 |
|
| 713 |
/* compute 1/x as double-double */ |
| 714 |
let yh = 1.0 / x; |
| 715 |
/* Assume 1 <= x < 2, then 0.5 <= yh <= 1, |
| 716 |
and yh = 1/x + eps with |eps| <= 2^-53. */ |
| 717 |
/* Newton's iteration for 1/x is y -> y + y*(1-x*y) */ |
| 718 |
let yl = yh * dd_fmla(-x, yh, 1.0); |
| 719 |
/* x*yh-1 = x*(1/x+eps)-1 = x*eps |
| 720 |
with |x*eps| <= 2^-52, thus the error on the FMA is bounded by |
| 721 |
ulp(2^-52.1) = 2^-105. |
| 722 |
Now |yl| <= |yh| * 2^-52 <= 2^-52, thus the rounding error on |
| 723 |
yh * __builtin_fma (-x, yh, 1.0) is bounded also by ulp(2^-52.1) = 2^-105. |
| 724 |
From [6], Lemma 3.7, if yl was computed exactly, then yh+yl would differ |
| 725 |
from 1/x by at most yh^2/theta^3*(1/x-yh)^2 for some theta in [yh,1/x] |
| 726 |
or [1/x,yh]. |
| 727 |
Since yh, 1/x <= 1, this gives eps^2 <= 2^-106. |
| 728 |
Adding the rounding errors, we have: |
| 729 |
|yh + yl - 1/x| <= 2^-105 + 2^-105 + 2^-106 < 2^-103.67. |
| 730 |
For the relative error, since |yh| >= 1/2, this gives: |
| 731 |
|yh + yl - 1/x| < 2^-102.67 * |yh+yl| |
| 732 |
*/ |
| 733 |
const THRESHOLD: [u64; 6] = [ |
| 734 |
0x3fbd500000000000, |
| 735 |
0x3fc59da6ca291ba6, |
| 736 |
0x3fcbc00000000000, |
| 737 |
0x3fd0c00000000000, |
| 738 |
0x3fd3800000000000, |
| 739 |
0x3fd6300000000000, |
| 740 |
]; |
| 741 |
let mut i = 0usize; |
| 742 |
while i < THRESHOLD.len() && yh > f64::from_bits(THRESHOLD[i]) { |
| 743 |
i += 1; |
| 744 |
} |
| 745 |
let p = ASYMPTOTIC_POLY[i]; |
| 746 |
u = DoubleDouble::from_exact_mult(yh, yh); |
| 747 |
/* Since |yh| <= 1, we have |uh| <= 1 and |ul| <= 2^-53. */ |
| 748 |
u.lo = dd_fmla(2.0 * yh, yl, u.lo); |
| 749 |
/* uh+ul approximates (yh+yl)^2, with absolute error bounded by |
| 750 |
ulp(ul) + yl^2, where ulp(ul) is the maximal rounding error in |
| 751 |
the FMA, and yl^2 is the neglected term. |
| 752 |
Since |ul| <= 2^-53, ulp(ul) <= 2^-105, and since |yl| <= 2^-52, |
| 753 |
this yields |uh + ul - yh^2| <= 2^-105 + 2^-104 < 2^-103.41. |
| 754 |
For the relative error, since |(yh+yl)^2| >= 1/4: |
| 755 |
|uh + ul - yh^2| < 2^-101.41 * |uh+ul|. |
| 756 |
And relatively to 1/x^2: |
| 757 |
yh + yl = 1/x * (1 + eps1) with |eps1| < 2^-102.67 |
| 758 |
uh + ul = (yh+yl)^2 * (1 + eps2) with |eps2| < 2^-101.41 |
| 759 |
This yields: |
| 760 |
|uh + ul - 1/x| < 2^-100.90 * |uh+ul|. |
| 761 |
*/ |
| 762 |
|
| 763 |
/* evaluate p(uh+ul) */ |
| 764 |
let mut zh: f64 = f64::from_bits(p[12]); // degree 23 |
| 765 |
zh = dd_fmla(zh, u.hi, f64::from_bits(p[11])); // degree 21 |
| 766 |
zh = dd_fmla(zh, u.hi, f64::from_bits(p[10])); // degree 19 |
| 767 |
|
| 768 |
/* degree 17: zh*(uh+ul)+p[i] */ |
| 769 |
let mut v = DoubleDouble::quick_f64_mult(zh, u); |
| 770 |
let mut z_dd = DoubleDouble::from_exact_add(f64::from_bits(p[9]), v.hi); |
| 771 |
z_dd.lo += v.lo; |
| 772 |
|
| 773 |
for a in (3..=15).rev().step_by(2) { |
| 774 |
v = DoubleDouble::quick_mult(z_dd, u); |
| 775 |
z_dd = DoubleDouble::from_exact_add(f64::from_bits(p[((a + 1) / 2) as usize]), v.hi); |
| 776 |
z_dd.lo += v.lo; |
| 777 |
} |
| 778 |
|
| 779 |
/* degree 1: (zh+zl)*(uh+ul)+p[0]+p[1] */ |
| 780 |
v = DoubleDouble::quick_mult(z_dd, u); |
| 781 |
z_dd = DoubleDouble::from_exact_add(f64::from_bits(p[0]), v.hi); |
| 782 |
z_dd.lo += v.lo + f64::from_bits(p[1]); |
| 783 |
/* multiply by yh+yl */ |
| 784 |
u = DoubleDouble::quick_mult(z_dd, DoubleDouble::new(yl, yh)); |
| 785 |
/* now uh+ul approximates p(1/x) */ |
| 786 |
/* now multiply (uh+ul)*(eh+el) */ |
| 787 |
v = DoubleDouble::quick_mult(u, e_dd); |
| 788 |
|
| 789 |
/* Write y = 1/x. We have the following errors: |
| 790 |
* the maximal mathematical error is: |
| 791 |
|erfc(x)*exp(x^2) - p(y)| < 2^-71.804 * |p(y)| (for i=3) thus |
| 792 |
|erfc(x) - exp(-x^2)*p(y)| < 2^-71.804 * |exp(-x^2)*p(y)| |
| 793 |
* the error in approximating exp(-x^2) by eh+el: |
| 794 |
|eh + el - exp(-x^2)| < 2^-74.139 * |eh + el| |
| 795 |
* the fact that we evaluate p on yh+yl instead of 1/x |
| 796 |
this error is bounded by |p'| * |yh+yl - 1/x|, where |
| 797 |
|yh+yl - 1/x| < 2^-102.67 * |yh+yl|, and the relative |
| 798 |
error is bounded by |p'/p| * |yh+yl - 1/x|. |
| 799 |
Since the maximal value of |p'/p| is bounded by 27.2 (for i=0), |
| 800 |
this yields 27.2 * 2^-102.67 < 2^-97.9 |
| 801 |
* the rounding errors when evaluating p on yh+yl: this error is bounded |
| 802 |
(relatively) by 2^-67.184 (for i=5), see analyze_erfc_asympt_fast() |
| 803 |
in erfc.sage |
| 804 |
* the rounding error in (uh+ul)*(eh+el): we assume this error is bounded |
| 805 |
by 2^-80 (relatively) |
| 806 |
This yields a global relative bound of: |
| 807 |
(1+2^-71.804)*(1+2^-74.139)*(1+2^-97.9)*(1+2^-67.184)*(1+2^-80)-1 |
| 808 |
< 2^-67.115 |
| 809 |
*/ |
| 810 |
if v.hi >= f64::from_bits(0x044151b9a3fdd5c9) { |
| 811 |
Erf { |
| 812 |
err: f64::from_bits(0x3bbd900000000000) * v.hi, |
| 813 |
result: v, |
| 814 |
} /* 2^-67.115 < 0x1.d9p-68 */ |
| 815 |
} else { |
| 816 |
Erf { |
| 817 |
result: v, |
| 818 |
err: f64::from_bits(0x0010000000000000), |
| 819 |
} // this overestimates 0x1.d9p-68 * h |
| 820 |
} |
| 821 |
} |
| 822 |
|
| 823 |
#[inline] |
| 824 |
fn erfc_fast(x: f64) -> Erf { |
| 825 |
if x < 0. |
| 826 |
// erfc(x) = 1 - erf(x) = 1 + erf(-x) |
| 827 |
{ |
| 828 |
let res = erf_fast(-x); |
| 829 |
/* h+l approximates erf(-x), with relative error bounded by err, |
| 830 |
where err <= 0x1.78p-69 */ |
| 831 |
let err = res.err * res.result.hi; /* convert into absolute error */ |
| 832 |
let mut t = DoubleDouble::from_exact_add(1.0, res.result.hi); |
| 833 |
t.lo += res.result.lo; |
| 834 |
// since h <= 2, the fast_two_sum() error is bounded by 2^-105*h <= 2^-104 |
| 835 |
/* After the fast_two_sum() call, we have |t| <= ulp(h) <= ulp(2) = 2^-51 |
| 836 |
thus assuming |l| <= 2^-51 after the cr_erf_fast() call, |
| 837 |
we have |t| <= 2^-50 here, thus the rounding |
| 838 |
error on t -= *l is bounded by ulp(2^-50) = 2^-102. |
| 839 |
The absolute error is thus bounded by err + 2^-104 + 2^-102 |
| 840 |
= err + 0x1.4p-102. |
| 841 |
The maximal value of err here is for |x| < 0.0625, where cr_erf_fast() |
| 842 |
returns 0x1.78p-69, and h=1/2, yielding err = 0x1.78p-70 here. |
| 843 |
Adding 0x1.4p-102 is thus exact. */ |
| 844 |
return Erf { |
| 845 |
err: err + f64::from_bits(0x3994000000000000), |
| 846 |
result: t, |
| 847 |
}; |
| 848 |
} else if x <= f64::from_bits(0x400713786d9c7c09) { |
| 849 |
let res = erf_fast(x); |
| 850 |
/* h+l approximates erf(x), with relative error bounded by err, |
| 851 |
where err <= 0x1.78p-69 */ |
| 852 |
let err = res.err * res.result.hi; /* convert into absolute error */ |
| 853 |
let mut t = DoubleDouble::from_exact_add(1.0, -res.result.hi); |
| 854 |
t.lo -= res.result.lo; |
| 855 |
/* for x >= 0x1.e861fbb24c00ap-2, erf(x) >= 1/2, thus 1-h is exact |
| 856 |
by Sterbenz theorem, thus t = 0 in fast_two_sum(), and we have t = -l |
| 857 |
here, thus the absolute error is err */ |
| 858 |
if x >= f64::from_bits(0x3fde861fbb24c00a) { |
| 859 |
return Erf { err, result: t }; |
| 860 |
} |
| 861 |
/* for x < 0x1.e861fbb24c00ap-2, the error in fast_two_sum() is bounded |
| 862 |
by 2^-105*h, and since h <= 1/2, this yields 2^-106. |
| 863 |
After the fast_two_sum() call, we have |t| <= ulp(h) <= ulp(1/2) = 2^-53 |
| 864 |
thus assuming |l| <= 2^-53 after the cr_erf_fast() call, |
| 865 |
we have |t| <= 2^-52 here, thus the rounding |
| 866 |
error on t -= *l is bounded by ulp(2^-52) = 2^-104. |
| 867 |
The absolute error is thus bounded by err + 2^-106 + 2^-104 |
| 868 |
The maximal value of err here is for x < 0.0625, where cr_erf_fast() |
| 869 |
returns 0x1.78p-69, and h=1/2, yielding err = 0x1.78p-70 here. |
| 870 |
Adding 0x1.4p-104 is thus exact. */ |
| 871 |
return Erf { |
| 872 |
err: err + f64::from_bits(0x3974000000000000), |
| 873 |
result: t, |
| 874 |
}; |
| 875 |
} |
| 876 |
/* Now THRESHOLD1 < x < 0x1.b39dc41e48bfdp+4 thus erfc(x) < 0.000046. */ |
| 877 |
/* on a i7-8700 with gcc 12.2.0, for x in [THRESHOLD1,+5.0], |
| 878 |
the average reciprocal throughput is about 111 cycles |
| 879 |
(among which 20 cycles for exp_1) */ |
| 880 |
erfc_asympt_fast(x) |
| 881 |
} |
| 882 |
|
| 883 |
/// Complementary error function |
| 884 |
/// |
| 885 |
/// Max ulp 0.5 |
| 886 |
pub fn f_erfc(x: f64) -> f64 { |
| 887 |
let t: u64 = x.to_bits(); |
| 888 |
let at: u64 = t & 0x7fff_ffff_ffff_ffff; |
| 889 |
|
| 890 |
if t >= 0x8000000000000000u64 |
| 891 |
// x = -NaN or x <= 0 (excluding +0) |
| 892 |
{ |
| 893 |
// for x <= -0x1.7744f8f74e94bp2, erfc(x) rounds to 2 (to nearest) |
| 894 |
if t >= 0xc017744f8f74e94bu64 |
| 895 |
// x = NaN or x <= -0x1.7744f8f74e94bp2 |
| 896 |
{ |
| 897 |
if t >= 0xfff0000000000000u64 { |
| 898 |
// -Inf or NaN |
| 899 |
if t == 0xfff0000000000000u64 { |
| 900 |
return 2.0; |
| 901 |
} // -Inf |
| 902 |
return x + x; // NaN |
| 903 |
} |
| 904 |
return black_box(2.0) - black_box(f64::from_bits(0x3c90000000000000)); // rounds to 2 or below(2) |
| 905 |
} |
| 906 |
|
| 907 |
// for -9.8390953768041405e-17 <= x <= 0, erfc(x) rounds to 1 (to nearest) |
| 908 |
if f64::from_bits(0xbc9c5bf891b4ef6a) <= x { |
| 909 |
return dd_fmla(-x, f64::from_bits(0x3c90000000000000), 1.0); |
| 910 |
} |
| 911 |
} else |
| 912 |
// x = +NaN or x >= 0 (excluding -0) |
| 913 |
{ |
| 914 |
// for x >= 0x1.b39dc41e48bfdp+4, erfc(x) < 2^-1075: rounds to 0 or 2^-1074 |
| 915 |
if at >= 0x403b39dc41e48bfdu64 |
| 916 |
// x = NaN or x >= 0x1.b39dc41e48bfdp+4 |
| 917 |
{ |
| 918 |
if at >= 0x7ff0000000000000u64 { |
| 919 |
// +Inf or NaN |
| 920 |
if at == 0x7ff0000000000000u64 { |
| 921 |
return 0.0; |
| 922 |
} // +Inf |
| 923 |
return x + x; // NaN |
| 924 |
} |
| 925 |
return black_box(f64::from_bits(0x0000000000000001)) * black_box(0.25); // 0 or 2^-1074 wrt rounding |
| 926 |
} |
| 927 |
|
| 928 |
// for 0 <= x <= 0x1.c5bf891b4ef6ap-55, erfc(x) rounds to 1 (to nearest) |
| 929 |
if x <= f64::from_bits(0x3c8c5bf891b4ef6a) { |
| 930 |
return dd_fmla(-x, f64::from_bits(0x3c90000000000000), 1.0); |
| 931 |
} |
| 932 |
} |
| 933 |
|
| 934 |
/* now -0x1.7744f8f74e94bp+2 < x < -0x1.c5bf891b4ef6ap-54 |
| 935 |
or 0x1.c5bf891b4ef6ap-55 < x < 0x1.b39dc41e48bfdp+4 */ |
| 936 |
|
| 937 |
let result = erfc_fast(x); |
| 938 |
|
| 939 |
let left = result.result.hi + (result.result.lo - result.err); |
| 940 |
let right = result.result.hi + (result.result.lo + result.err); |
| 941 |
if left == right { |
| 942 |
return left; |
| 943 |
} |
| 944 |
erfc_accurate(x) |
| 945 |
} |
| 946 |
|
| 947 |
#[cfg(test)] |
| 948 |
mod tests { |
| 949 |
use super::*; |
| 950 |
#[test] |
| 951 |
fn test_erfc() { |
| 952 |
assert_eq!(f_erfc(1.0), 0.15729920705028513); |
| 953 |
assert_eq!(f_erfc(0.5), 0.4795001221869535); |
| 954 |
assert_eq!(f_erfc(0.000000005), 0.9999999943581042); |
| 955 |
assert_eq!(f_erfc(-0.00000000000065465465423305), 1.0000000000007387); |
| 956 |
assert!(f_erfc(f64::NAN).is_nan()); |
| 957 |
assert_eq!(f_erfc(f64::INFINITY), 0.0); |
| 958 |
assert_eq!(f_erfc(f64::NEG_INFINITY), 2.0); |
| 959 |
} |
| 960 |
} |