| 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::bessel::alpha0::{ |
| 30 |
bessel_0_asympt_alpha, bessel_0_asympt_alpha_fast, bessel_0_asympt_alpha_hard, |
| 31 |
}; |
| 32 |
use crate::bessel::beta0::{ |
| 33 |
bessel_0_asympt_beta, bessel_0_asympt_beta_fast, bessel_0_asympt_beta_hard, |
| 34 |
}; |
| 35 |
use crate::bessel::i0::bessel_rsqrt_hard; |
| 36 |
use crate::bessel::j0_coeffs_remez::J0_COEFFS_REMEZ; |
| 37 |
use crate::bessel::j0_coeffs_taylor::J0_COEFFS_TAYLOR; |
| 38 |
use crate::bessel::j0f_coeffs::{J0_ZEROS, J0_ZEROS_VALUE}; |
| 39 |
use crate::common::f_fmla; |
| 40 |
use crate::double_double::DoubleDouble; |
| 41 |
use crate::dyadic_float::{DyadicFloat128, DyadicSign}; |
| 42 |
use crate::polyeval::{f_polyeval9, f_polyeval10, f_polyeval12, f_polyeval19}; |
| 43 |
use crate::rounding::CpuCeil; |
| 44 |
use crate::sin_helper::{cos_dd_small, cos_dd_small_fast, cos_f128_small}; |
| 45 |
use crate::sincos_reduce::{AngleReduced, rem2pi_any, rem2pi_f128}; |
| 46 |
|
| 47 |
/// Bessel of the first kind of order 0 |
| 48 |
pub fn f_j0(x: f64) -> f64 { |
| 49 |
let ux = x.to_bits().wrapping_shl(1); |
| 50 |
|
| 51 |
if ux >= 0x7ffu64 << 53 || ux <= 0x7960000000000000u64 { |
| 52 |
// |x| <= f64::EPSILON, |x| == inf, |x| == NaN |
| 53 |
if ux <= 0x77723ef88126da90u64 { |
| 54 |
// |x| <= 0.00000000000000000000532 |
| 55 |
return 1.; |
| 56 |
} |
| 57 |
if ux <= 0x7960000000000000u64 { |
| 58 |
// |x| <= f64::EPSILON |
| 59 |
// J0(x) ~ 1-x^2/4+O[x]^4 |
| 60 |
let half_x = 0.5 * x; // exact. |
| 61 |
return f_fmla(half_x, -half_x, 1.); |
| 62 |
} |
| 63 |
if x.is_infinite() { |
| 64 |
return 0.; |
| 65 |
} |
| 66 |
return x + f64::NAN; // x == NaN |
| 67 |
} |
| 68 |
|
| 69 |
let x_abs = x.to_bits() & 0x7fff_ffff_ffff_ffff; |
| 70 |
|
| 71 |
let ax = f64::from_bits(x_abs); |
| 72 |
|
| 73 |
if x_abs <= 0x4052b33333333333u64 { |
| 74 |
// |x| <= 74.8 |
| 75 |
if x_abs <= 0x3ff199999999999au64 { |
| 76 |
// |x| <= 1.1 |
| 77 |
return j0_maclaurin_series_fast(ax); |
| 78 |
} |
| 79 |
return j0_small_argument_fast(ax); |
| 80 |
} |
| 81 |
|
| 82 |
j0_asympt_fast(ax) |
| 83 |
} |
| 84 |
|
| 85 |
/** |
| 86 |
Generated by SageMath: |
| 87 |
```python |
| 88 |
mp.prec = 180 |
| 89 |
def print_expansion_at_0(): |
| 90 |
print(f"const J0_MACLAURIN_SERIES: [(u64, u64); 12] = [") |
| 91 |
from mpmath import mp, j0, taylor |
| 92 |
poly = taylor(lambda val: j0(val), 0, 24) |
| 93 |
real_i = 0 |
| 94 |
for i in range(0, 24, 2): |
| 95 |
print_double_double("", DD(poly[i])) |
| 96 |
real_i = real_i + 1 |
| 97 |
print("];") |
| 98 |
print(poly) |
| 99 |
|
| 100 |
print_expansion_at_0() |
| 101 |
``` |
| 102 |
**/ |
| 103 |
#[inline] |
| 104 |
pub(crate) fn j0_maclaurin_series_fast(x: f64) -> f64 { |
| 105 |
const C: [u64; 12] = [ |
| 106 |
0x3ff0000000000000, |
| 107 |
0xbfd0000000000000, |
| 108 |
0x3f90000000000000, |
| 109 |
0xbf3c71c71c71c71c, |
| 110 |
0x3edc71c71c71c71c, |
| 111 |
0xbe723456789abcdf, |
| 112 |
0x3e002e85c0898b71, |
| 113 |
0xbd8522a43f65486a, |
| 114 |
0x3d0522a43f65486a, |
| 115 |
0xbc80b313289be0b9, |
| 116 |
0x3bf5601885e63e5d, |
| 117 |
0xbb669ca9cf3b7f54, |
| 118 |
]; |
| 119 |
|
| 120 |
let dx2 = DoubleDouble::from_exact_mult(x, x); |
| 121 |
|
| 122 |
let p = f_polyeval10( |
| 123 |
dx2.hi, |
| 124 |
f64::from_bits(C[2]), |
| 125 |
f64::from_bits(C[3]), |
| 126 |
f64::from_bits(C[4]), |
| 127 |
f64::from_bits(C[5]), |
| 128 |
f64::from_bits(C[6]), |
| 129 |
f64::from_bits(C[7]), |
| 130 |
f64::from_bits(C[8]), |
| 131 |
f64::from_bits(C[9]), |
| 132 |
f64::from_bits(C[10]), |
| 133 |
f64::from_bits(C[11]), |
| 134 |
); |
| 135 |
let mut z = DoubleDouble::mul_f64_add_f64(dx2, p, f64::from_bits(C[1])); |
| 136 |
z = DoubleDouble::mul_add_f64(dx2, z, f64::from_bits(C[0])); |
| 137 |
|
| 138 |
// squaring error (2^-56) + poly error 2^-75 |
| 139 |
let err = f_fmla( |
| 140 |
dx2.hi, |
| 141 |
f64::from_bits(0x3c70000000000000), // 2^-56 |
| 142 |
f64::from_bits(0x3b40000000000000), // 2^-75 |
| 143 |
); |
| 144 |
let ub = z.hi + (z.lo + err); |
| 145 |
let lb = z.hi + (z.lo - err); |
| 146 |
|
| 147 |
if ub == lb { |
| 148 |
return z.to_f64(); |
| 149 |
} |
| 150 |
j0_maclaurin_series(x) |
| 151 |
} |
| 152 |
|
| 153 |
/** |
| 154 |
Generated by SageMath: |
| 155 |
```python |
| 156 |
mp.prec = 180 |
| 157 |
def print_expansion_at_0(): |
| 158 |
print(f"const J0_MACLAURIN_SERIES: [(u64, u64); 12] = [") |
| 159 |
from mpmath import mp, j0, taylor |
| 160 |
poly = taylor(lambda val: j0(val), 0, 24) |
| 161 |
real_i = 0 |
| 162 |
for i in range(0, 24, 2): |
| 163 |
print_double_double("", DD(poly[i])) |
| 164 |
real_i = real_i + 1 |
| 165 |
print("];") |
| 166 |
print(poly) |
| 167 |
|
| 168 |
print_expansion_at_0() |
| 169 |
``` |
| 170 |
**/ |
| 171 |
#[cold] |
| 172 |
pub(crate) fn j0_maclaurin_series(x: f64) -> f64 { |
| 173 |
const C: [(u64, u64); 12] = [ |
| 174 |
(0x0000000000000000, 0x3ff0000000000000), |
| 175 |
(0x0000000000000000, 0xbfd0000000000000), |
| 176 |
(0x0000000000000000, 0x3f90000000000000), |
| 177 |
(0xbbdc71c71c71c71c, 0xbf3c71c71c71c71c), |
| 178 |
(0x3b7c71c71c71c71c, 0x3edc71c71c71c71c), |
| 179 |
(0xbab23456789abcdf, 0xbe723456789abcdf), |
| 180 |
(0xba8b6edec0692e65, 0x3e002e85c0898b71), |
| 181 |
(0x3a2604db055bd075, 0xbd8522a43f65486a), |
| 182 |
(0xb9a604db055bd075, 0x3d0522a43f65486a), |
| 183 |
(0x3928824198c6f6e1, 0xbc80b313289be0b9), |
| 184 |
(0xb869b0b430eb27b8, 0x3bf5601885e63e5d), |
| 185 |
(0x380ee6b4638f3a25, 0xbb669ca9cf3b7f54), |
| 186 |
]; |
| 187 |
|
| 188 |
let dx2 = DoubleDouble::from_exact_mult(x, x); |
| 189 |
|
| 190 |
let p = f_polyeval12( |
| 191 |
dx2, |
| 192 |
DoubleDouble::from_bit_pair(C[0]), |
| 193 |
DoubleDouble::from_bit_pair(C[1]), |
| 194 |
DoubleDouble::from_bit_pair(C[2]), |
| 195 |
DoubleDouble::from_bit_pair(C[3]), |
| 196 |
DoubleDouble::from_bit_pair(C[4]), |
| 197 |
DoubleDouble::from_bit_pair(C[5]), |
| 198 |
DoubleDouble::from_bit_pair(C[6]), |
| 199 |
DoubleDouble::from_bit_pair(C[7]), |
| 200 |
DoubleDouble::from_bit_pair(C[8]), |
| 201 |
DoubleDouble::from_bit_pair(C[9]), |
| 202 |
DoubleDouble::from_bit_pair(C[10]), |
| 203 |
DoubleDouble::from_bit_pair(C[11]), |
| 204 |
); |
| 205 |
let r = DoubleDouble::from_exact_add(p.hi, p.lo); |
| 206 |
const ERR: f64 = f64::from_bits(0x39d0000000000000); // 2^-98 |
| 207 |
let ub = r.hi + (r.lo + ERR); |
| 208 |
let lb = r.hi + (r.lo - ERR); |
| 209 |
if ub == lb { |
| 210 |
return r.to_f64(); |
| 211 |
} |
| 212 |
j0_maclaurin_series_hard(x) |
| 213 |
} |
| 214 |
|
| 215 |
/** |
| 216 |
Generated by SageMath: |
| 217 |
|
| 218 |
```python |
| 219 |
mp.prec = 180 |
| 220 |
def print_expansion_at_0(): |
| 221 |
print(f"const P: [DyadicFloat128; 12] = [") |
| 222 |
from mpmath import mp, j0, taylor |
| 223 |
poly = taylor(lambda val: j0(val), 0, 24) |
| 224 |
# print(poly) |
| 225 |
real_i = 0 |
| 226 |
for i in range(0, 24, 2): |
| 227 |
print_dyadic(DD(poly[i])) |
| 228 |
real_i = real_i + 1 |
| 229 |
print("];") |
| 230 |
print(poly) |
| 231 |
|
| 232 |
print_expansion_at_0() |
| 233 |
``` |
| 234 |
**/ |
| 235 |
#[cold] |
| 236 |
#[inline(never)] |
| 237 |
pub(crate) fn j0_maclaurin_series_hard(x: f64) -> f64 { |
| 238 |
static P: [DyadicFloat128; 12] = [ |
| 239 |
DyadicFloat128 { |
| 240 |
sign: DyadicSign::Pos, |
| 241 |
exponent: -127, |
| 242 |
mantissa: 0x80000000_00000000_00000000_00000000_u128, |
| 243 |
}, |
| 244 |
DyadicFloat128 { |
| 245 |
sign: DyadicSign::Neg, |
| 246 |
exponent: -129, |
| 247 |
mantissa: 0x80000000_00000000_00000000_00000000_u128, |
| 248 |
}, |
| 249 |
DyadicFloat128 { |
| 250 |
sign: DyadicSign::Pos, |
| 251 |
exponent: -133, |
| 252 |
mantissa: 0x80000000_00000000_00000000_00000000_u128, |
| 253 |
}, |
| 254 |
DyadicFloat128 { |
| 255 |
sign: DyadicSign::Neg, |
| 256 |
exponent: -139, |
| 257 |
mantissa: 0xe38e38e3_8e38e38e_38e38e38_e38e38e4_u128, |
| 258 |
}, |
| 259 |
DyadicFloat128 { |
| 260 |
sign: DyadicSign::Pos, |
| 261 |
exponent: -145, |
| 262 |
mantissa: 0xe38e38e3_8e38e38e_38e38e38_e38e38e4_u128, |
| 263 |
}, |
| 264 |
DyadicFloat128 { |
| 265 |
sign: DyadicSign::Neg, |
| 266 |
exponent: -151, |
| 267 |
mantissa: 0x91a2b3c4_d5e6f809_1a2b3c4d_5e6f8092_u128, |
| 268 |
}, |
| 269 |
DyadicFloat128 { |
| 270 |
sign: DyadicSign::Pos, |
| 271 |
exponent: -158, |
| 272 |
mantissa: 0x81742e04_4c5b8724_8909fcb6_8cd4e410_u128, |
| 273 |
}, |
| 274 |
DyadicFloat128 { |
| 275 |
sign: DyadicSign::Neg, |
| 276 |
exponent: -166, |
| 277 |
mantissa: 0xa91521fb_2a434d3f_649f5485_f169a743_u128, |
| 278 |
}, |
| 279 |
DyadicFloat128 { |
| 280 |
sign: DyadicSign::Pos, |
| 281 |
exponent: -174, |
| 282 |
mantissa: 0xa91521fb_2a434d3f_649f5485_f169a743_u128, |
| 283 |
}, |
| 284 |
DyadicFloat128 { |
| 285 |
sign: DyadicSign::Neg, |
| 286 |
exponent: -182, |
| 287 |
mantissa: 0x85989944_df05c4ef_b7cce721_23e1b391_u128, |
| 288 |
}, |
| 289 |
DyadicFloat128 { |
| 290 |
sign: DyadicSign::Pos, |
| 291 |
exponent: -191, |
| 292 |
mantissa: 0xab00c42f_31f2e799_3d2f3c53_6120e5d8_u128, |
| 293 |
}, |
| 294 |
DyadicFloat128 { |
| 295 |
sign: DyadicSign::Neg, |
| 296 |
exponent: -200, |
| 297 |
mantissa: 0xb4e54e79_dbfa9c23_29738e18_bb602809_u128, |
| 298 |
}, |
| 299 |
]; |
| 300 |
let dx = DyadicFloat128::new_from_f64(x); |
| 301 |
let x2 = dx * dx; |
| 302 |
|
| 303 |
let mut p = P[11]; |
| 304 |
for i in (0..11).rev() { |
| 305 |
p = x2 * p + P[i]; |
| 306 |
} |
| 307 |
|
| 308 |
p.fast_as_f64() |
| 309 |
} |
| 310 |
|
| 311 |
/// This method on small range searches for nearest zero or extremum. |
| 312 |
/// Then picks stored series expansion at the point end evaluates the poly at the point. |
| 313 |
#[inline] |
| 314 |
pub(crate) fn j0_small_argument_fast(x: f64) -> f64 { |
| 315 |
// let avg_step = 74.6145 / 47.0; |
| 316 |
// let inv_step = 1.0 / avg_step; |
| 317 |
|
| 318 |
const INV_STEP: f64 = 0.6299043751549631; |
| 319 |
|
| 320 |
let fx = x * INV_STEP; |
| 321 |
const J0_ZEROS_COUNT: f64 = (J0_ZEROS.len() - 1) as f64; |
| 322 |
let idx0 = unsafe { fx.min(J0_ZEROS_COUNT).to_int_unchecked::<usize>() }; |
| 323 |
let idx1 = unsafe { |
| 324 |
fx.cpu_ceil() |
| 325 |
.min(J0_ZEROS_COUNT) |
| 326 |
.to_int_unchecked::<usize>() |
| 327 |
}; |
| 328 |
|
| 329 |
let found_zero0 = DoubleDouble::from_bit_pair(J0_ZEROS[idx0]); |
| 330 |
let found_zero1 = DoubleDouble::from_bit_pair(J0_ZEROS[idx1]); |
| 331 |
|
| 332 |
let dist0 = (found_zero0.hi - x).abs(); |
| 333 |
let dist1 = (found_zero1.hi - x).abs(); |
| 334 |
|
| 335 |
let (found_zero, idx, dist) = if dist0 < dist1 { |
| 336 |
(found_zero0, idx0, dist0) |
| 337 |
} else { |
| 338 |
(found_zero1, idx1, dist1) |
| 339 |
}; |
| 340 |
|
| 341 |
if idx == 0 { |
| 342 |
return j0_maclaurin_series_fast(x); |
| 343 |
} |
| 344 |
|
| 345 |
let is_too_close_too_zero = dist.abs() < 1e-3; |
| 346 |
|
| 347 |
let c = if is_too_close_too_zero { |
| 348 |
&J0_COEFFS_TAYLOR[idx - 1] |
| 349 |
} else { |
| 350 |
&J0_COEFFS_REMEZ[idx - 1] |
| 351 |
}; |
| 352 |
|
| 353 |
let r = DoubleDouble::full_add_f64(-found_zero, x.abs()); |
| 354 |
|
| 355 |
// We hit exact zero, value, better to return it directly |
| 356 |
if dist == 0. { |
| 357 |
return f64::from_bits(J0_ZEROS_VALUE[idx]); |
| 358 |
} |
| 359 |
let p = f_polyeval19( |
| 360 |
r.hi, |
| 361 |
f64::from_bits(c[5].1), |
| 362 |
f64::from_bits(c[6].1), |
| 363 |
f64::from_bits(c[7].1), |
| 364 |
f64::from_bits(c[8].1), |
| 365 |
f64::from_bits(c[9].1), |
| 366 |
f64::from_bits(c[10].1), |
| 367 |
f64::from_bits(c[11].1), |
| 368 |
f64::from_bits(c[12].1), |
| 369 |
f64::from_bits(c[13].1), |
| 370 |
f64::from_bits(c[14].1), |
| 371 |
f64::from_bits(c[15].1), |
| 372 |
f64::from_bits(c[16].1), |
| 373 |
f64::from_bits(c[17].1), |
| 374 |
f64::from_bits(c[18].1), |
| 375 |
f64::from_bits(c[19].1), |
| 376 |
f64::from_bits(c[20].1), |
| 377 |
f64::from_bits(c[21].1), |
| 378 |
f64::from_bits(c[22].1), |
| 379 |
f64::from_bits(c[23].1), |
| 380 |
); |
| 381 |
|
| 382 |
let mut z = DoubleDouble::mul_f64_add(r, p, DoubleDouble::from_bit_pair(c[4])); |
| 383 |
z = DoubleDouble::mul_add(z, r, DoubleDouble::from_bit_pair(c[3])); |
| 384 |
z = DoubleDouble::mul_add(z, r, DoubleDouble::from_bit_pair(c[2])); |
| 385 |
z = DoubleDouble::mul_add(z, r, DoubleDouble::from_bit_pair(c[1])); |
| 386 |
z = DoubleDouble::mul_add(z, r, DoubleDouble::from_bit_pair(c[0])); |
| 387 |
let err = f_fmla( |
| 388 |
z.hi, |
| 389 |
f64::from_bits(0x3c70000000000000), // 2^-56 |
| 390 |
f64::from_bits(0x3bf0000000000000), // 2^-64 |
| 391 |
); |
| 392 |
let ub = z.hi + (z.lo + err); |
| 393 |
let lb = z.hi + (z.lo - err); |
| 394 |
|
| 395 |
if ub == lb { |
| 396 |
return z.to_f64(); |
| 397 |
} |
| 398 |
|
| 399 |
j0_small_argument_dd(r, c) |
| 400 |
} |
| 401 |
|
| 402 |
#[cold] |
| 403 |
fn j0_small_argument_dd(r: DoubleDouble, c0: &[(u64, u64); 24]) -> f64 { |
| 404 |
let c = &c0[15..]; |
| 405 |
|
| 406 |
let p0 = f_polyeval9( |
| 407 |
r.to_f64(), |
| 408 |
f64::from_bits(c[0].1), |
| 409 |
f64::from_bits(c[1].1), |
| 410 |
f64::from_bits(c[2].1), |
| 411 |
f64::from_bits(c[3].1), |
| 412 |
f64::from_bits(c[4].1), |
| 413 |
f64::from_bits(c[5].1), |
| 414 |
f64::from_bits(c[6].1), |
| 415 |
f64::from_bits(c[7].1), |
| 416 |
f64::from_bits(c[8].1), |
| 417 |
); |
| 418 |
|
| 419 |
let c = c0; |
| 420 |
|
| 421 |
let mut p_e = DoubleDouble::mul_f64_add(r, p0, DoubleDouble::from_bit_pair(c[14])); |
| 422 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[13])); |
| 423 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[12])); |
| 424 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[11])); |
| 425 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[10])); |
| 426 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[9])); |
| 427 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[8])); |
| 428 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[7])); |
| 429 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[6])); |
| 430 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[5])); |
| 431 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[4])); |
| 432 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[3])); |
| 433 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[2])); |
| 434 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[1])); |
| 435 |
p_e = DoubleDouble::mul_add(p_e, r, DoubleDouble::from_bit_pair(c[0])); |
| 436 |
|
| 437 |
let p = DoubleDouble::from_exact_add(p_e.hi, p_e.lo); |
| 438 |
let err = f_fmla( |
| 439 |
p.hi, |
| 440 |
f64::from_bits(0x3c10000000000000), // 2^-62 |
| 441 |
f64::from_bits(0x3a90000000000000), // 2^-86 |
| 442 |
); |
| 443 |
let ub = p.hi + (p.lo + err); |
| 444 |
let lb = p.hi + (p.lo - err); |
| 445 |
if ub != lb { |
| 446 |
return j0_small_argument_hard(r, c); |
| 447 |
} |
| 448 |
p.to_f64() |
| 449 |
} |
| 450 |
|
| 451 |
#[cold] |
| 452 |
#[inline(never)] |
| 453 |
fn j0_small_argument_hard(r: DoubleDouble, c: &[(u64, u64); 24]) -> f64 { |
| 454 |
let mut p = DoubleDouble::from_bit_pair(c[23]); |
| 455 |
for i in (0..23).rev() { |
| 456 |
p = DoubleDouble::mul_add(r, p, DoubleDouble::from_bit_pair(c[i])); |
| 457 |
p = DoubleDouble::from_exact_add(p.hi, p.lo); |
| 458 |
} |
| 459 |
p.to_f64() |
| 460 |
} |
| 461 |
|
| 462 |
/* |
| 463 |
Evaluates: |
| 464 |
J0 = sqrt(2/(PI*x)) * beta(x) * cos(x - PI/4 - alpha(x)) |
| 465 |
*/ |
| 466 |
#[inline] |
| 467 |
pub(crate) fn j0_asympt_fast(x: f64) -> f64 { |
| 468 |
let x = x.abs(); |
| 469 |
|
| 470 |
const SQRT_2_OVER_PI: DoubleDouble = DoubleDouble::new( |
| 471 |
f64::from_bits(0xbc8cbc0d30ebfd15), |
| 472 |
f64::from_bits(0x3fe9884533d43651), |
| 473 |
); |
| 474 |
|
| 475 |
const MPI_OVER_4: DoubleDouble = DoubleDouble::new( |
| 476 |
f64::from_bits(0xbc81a62633145c07), |
| 477 |
f64::from_bits(0xbfe921fb54442d18), |
| 478 |
); |
| 479 |
|
| 480 |
let recip = if x.to_bits() > 0x7fd000000000000u64 { |
| 481 |
DoubleDouble::quick_mult_f64(DoubleDouble::from_exact_div_fma(4.0, x), 0.25) |
| 482 |
} else { |
| 483 |
DoubleDouble::from_recip(x) |
| 484 |
}; |
| 485 |
|
| 486 |
let alpha = bessel_0_asympt_alpha_fast(recip); |
| 487 |
let beta = bessel_0_asympt_beta_fast(recip); |
| 488 |
|
| 489 |
let AngleReduced { angle } = rem2pi_any(x); |
| 490 |
|
| 491 |
// Without full subtraction cancellation happens sometimes |
| 492 |
let x0pi34 = DoubleDouble::full_dd_sub(MPI_OVER_4, alpha); |
| 493 |
let r0 = DoubleDouble::full_dd_add(angle, x0pi34); |
| 494 |
|
| 495 |
let m_cos = cos_dd_small_fast(r0); |
| 496 |
let z0 = DoubleDouble::quick_mult(beta, m_cos); |
| 497 |
let r_sqrt = DoubleDouble::from_rsqrt_fast(x); |
| 498 |
let scale = DoubleDouble::quick_mult(SQRT_2_OVER_PI, r_sqrt); |
| 499 |
let p = DoubleDouble::quick_mult(scale, z0); |
| 500 |
|
| 501 |
let err = f_fmla( |
| 502 |
p.hi, |
| 503 |
f64::from_bits(0x3be0000000000000), // 2^-65 |
| 504 |
f64::from_bits(0x3a60000000000000), // 2^-89 |
| 505 |
); |
| 506 |
let ub = p.hi + (p.lo + err); |
| 507 |
let lb = p.hi + (p.lo - err); |
| 508 |
|
| 509 |
if ub == lb { |
| 510 |
return p.to_f64(); |
| 511 |
} |
| 512 |
j0_asympt(x, recip, r_sqrt, angle) |
| 513 |
} |
| 514 |
|
| 515 |
/* |
| 516 |
Evaluates: |
| 517 |
J0 = sqrt(2/(PI*x)) * beta(x) * cos(x - PI/4 - alpha(x)) |
| 518 |
*/ |
| 519 |
pub(crate) fn j0_asympt( |
| 520 |
x: f64, |
| 521 |
recip: DoubleDouble, |
| 522 |
r_sqrt: DoubleDouble, |
| 523 |
angle: DoubleDouble, |
| 524 |
) -> f64 { |
| 525 |
const SQRT_2_OVER_PI: DoubleDouble = DoubleDouble::new( |
| 526 |
f64::from_bits(0xbc8cbc0d30ebfd15), |
| 527 |
f64::from_bits(0x3fe9884533d43651), |
| 528 |
); |
| 529 |
|
| 530 |
const MPI_OVER_4: DoubleDouble = DoubleDouble::new( |
| 531 |
f64::from_bits(0xbc81a62633145c07), |
| 532 |
f64::from_bits(0xbfe921fb54442d18), |
| 533 |
); |
| 534 |
|
| 535 |
let alpha = bessel_0_asympt_alpha(recip); |
| 536 |
let beta = bessel_0_asympt_beta(recip); |
| 537 |
|
| 538 |
// Without full subtraction cancellation happens sometimes |
| 539 |
let x0pi34 = DoubleDouble::full_dd_sub(MPI_OVER_4, alpha); |
| 540 |
let r0 = DoubleDouble::full_dd_add(angle, x0pi34); |
| 541 |
|
| 542 |
let m_cos = cos_dd_small(r0); |
| 543 |
let z0 = DoubleDouble::quick_mult(beta, m_cos); |
| 544 |
let scale = DoubleDouble::quick_mult(SQRT_2_OVER_PI, r_sqrt); |
| 545 |
let r = DoubleDouble::quick_mult(scale, z0); |
| 546 |
|
| 547 |
let p = DoubleDouble::from_exact_add(r.hi, r.lo); |
| 548 |
let err = f_fmla( |
| 549 |
p.hi, |
| 550 |
f64::from_bits(0x3bd0000000000000), // 2^-66 |
| 551 |
f64::from_bits(0x39e0000000000000), // 2^-97 |
| 552 |
); |
| 553 |
let ub = p.hi + (p.lo + err); |
| 554 |
let lb = p.hi + (p.lo - err); |
| 555 |
|
| 556 |
if ub == lb { |
| 557 |
return p.to_f64(); |
| 558 |
} |
| 559 |
j0_asympt_hard(x) |
| 560 |
} |
| 561 |
|
| 562 |
/* |
| 563 |
Evaluates: |
| 564 |
J0 = sqrt(2/(PI*x)) * beta(x) * cos(x - PI/4 - alpha(x)) |
| 565 |
*/ |
| 566 |
#[cold] |
| 567 |
#[inline(never)] |
| 568 |
pub(crate) fn j0_asympt_hard(x: f64) -> f64 { |
| 569 |
const SQRT_2_OVER_PI: DyadicFloat128 = DyadicFloat128 { |
| 570 |
sign: DyadicSign::Pos, |
| 571 |
exponent: -128, |
| 572 |
mantissa: 0xcc42299e_a1b28468_7e59e280_5d5c7180_u128, |
| 573 |
}; |
| 574 |
|
| 575 |
const MPI_OVER_4: DyadicFloat128 = DyadicFloat128 { |
| 576 |
sign: DyadicSign::Neg, |
| 577 |
exponent: -128, |
| 578 |
mantissa: 0xc90fdaa2_2168c234_c4c6628b_80dc1cd1_u128, |
| 579 |
}; |
| 580 |
|
| 581 |
let x_dyadic = DyadicFloat128::new_from_f64(x); |
| 582 |
let recip = DyadicFloat128::accurate_reciprocal(x); |
| 583 |
|
| 584 |
let alpha = bessel_0_asympt_alpha_hard(recip); |
| 585 |
let beta = bessel_0_asympt_beta_hard(recip); |
| 586 |
|
| 587 |
let angle = rem2pi_f128(x_dyadic); |
| 588 |
|
| 589 |
let x0pi34 = MPI_OVER_4 - alpha; |
| 590 |
let r0 = angle + x0pi34; |
| 591 |
|
| 592 |
let m_sin = cos_f128_small(r0); |
| 593 |
|
| 594 |
let z0 = beta * m_sin; |
| 595 |
let r_sqrt = bessel_rsqrt_hard(x, recip); |
| 596 |
let scale = SQRT_2_OVER_PI * r_sqrt; |
| 597 |
let p = scale * z0; |
| 598 |
p.fast_as_f64() |
| 599 |
} |
| 600 |
|
| 601 |
#[cfg(test)] |
| 602 |
mod tests { |
| 603 |
use super::*; |
| 604 |
|
| 605 |
#[test] |
| 606 |
fn test_j0() { |
| 607 |
assert_eq!(f_j0(f64::EPSILON), 1.0); |
| 608 |
assert_eq!(f_j0(-0.000000000000000000000532), 1.0); |
| 609 |
assert_eq!(f_j0(0.0000000000000000000532), 1.0); |
| 610 |
assert_eq!(f_j0(-2.000976555054876), 0.22332760641907712); |
| 611 |
assert_eq!(f_j0(-2.3369499004222215E+304), -3.3630754230844632e-155); |
| 612 |
assert_eq!( |
| 613 |
f_j0(f64::from_bits(0xd71a31ffe2ff7e9f)), |
| 614 |
f64::from_bits(0xb2e58532f95056ff) |
| 615 |
); |
| 616 |
assert_eq!(f_j0(6.1795701510782757E+307), 6.075192922402001e-155); |
| 617 |
assert_eq!(f_j0(6.1795701510782757E+301), 4.118334155030934e-152); |
| 618 |
assert_eq!(f_j0(6.1795701510782757E+157), 9.5371668900364e-80); |
| 619 |
assert_eq!(f_j0(79.), -0.08501719554953485); |
| 620 |
// Without FMA 2.703816901253004e-16 |
| 621 |
#[cfg(any( |
| 622 |
all(target_arch = "x86_64", target_feature = "fma"), |
| 623 |
target_arch = "aarch64" |
| 624 |
))] |
| 625 |
assert_eq!(f_j0(93.463718781944774171190), 2.7038169012530046e-16); |
| 626 |
assert_eq!(f_j0(99.746819858680596470279979), -8.419106281522749e-17); |
| 627 |
assert_eq!(f_j0(f64::INFINITY), 0.); |
| 628 |
assert_eq!(f_j0(f64::NEG_INFINITY), 0.); |
| 629 |
assert!(f_j0(f64::NAN).is_nan()); |
| 630 |
} |
| 631 |
} |