| 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, f_fmla}; |
| 30 |
use crate::double_double::DoubleDouble; |
| 31 |
use crate::hyperbolic::acosh::{ |
| 32 |
ACOSH_ASINH_B, ACOSH_ASINH_L1, ACOSH_ASINH_L2, ACOSH_ASINH_LL, ACOSH_ASINH_R1, ACOSH_ASINH_R2, |
| 33 |
ACOSH_ASINH_REFINE_T2, ACOSH_ASINH_REFINE_T4, ACOSH_SINH_REFINE_T1, ACOSH_SINH_REFINE_T3, |
| 34 |
lpoly_xd_generic, |
| 35 |
}; |
| 36 |
|
| 37 |
#[cold] |
| 38 |
fn asinh_refine(x: f64, a: f64, z: DoubleDouble) -> f64 { |
| 39 |
let mut t = z.hi.to_bits(); |
| 40 |
let ex: i32 = (t >> 52) as i32; |
| 41 |
let e = ex.wrapping_sub(0x3ff) + if z.lo == 0.0 { 1i32 } else { 0i32 }; |
| 42 |
t &= 0x000fffffffffffff; |
| 43 |
t |= 0x3ffu64 << 52; |
| 44 |
let ed = e as f64; |
| 45 |
let v = (a - ed + f64::from_bits(0x3ff0000800000000)).to_bits(); |
| 46 |
let i = (v.wrapping_sub(0x3ffu64 << 52)) >> (52 - 16); |
| 47 |
let i1 = (i >> 12) & 0x1f; |
| 48 |
let i2 = (i >> 8) & 0xf; |
| 49 |
let i3 = (i >> 4) & 0xf; |
| 50 |
let i4 = i & 0xf; |
| 51 |
const L20: f64 = f64::from_bits(0x3fd62e42fefa3800); |
| 52 |
const L21: f64 = f64::from_bits(0x3d1ef35793c76800); |
| 53 |
const L22: f64 = f64::from_bits(0xba49ff0342542fc3); |
| 54 |
let el2 = L22 * ed; |
| 55 |
let el1 = L21 * ed; |
| 56 |
let el0 = L20 * ed; |
| 57 |
let mut dl0: f64; |
| 58 |
|
| 59 |
let ll0i1 = ACOSH_ASINH_LL[0][i1 as usize]; |
| 60 |
let ll1i2 = ACOSH_ASINH_LL[1][i2 as usize]; |
| 61 |
let ll2i3 = ACOSH_ASINH_LL[2][i3 as usize]; |
| 62 |
let ll3i4 = ACOSH_ASINH_LL[3][i4 as usize]; |
| 63 |
|
| 64 |
dl0 = f64::from_bits(ll0i1.0) |
| 65 |
+ f64::from_bits(ll1i2.0) |
| 66 |
+ (f64::from_bits(ll2i3.0) + f64::from_bits(ll3i4.0)); |
| 67 |
let dl1 = f64::from_bits(ll0i1.1) |
| 68 |
+ f64::from_bits(ll1i2.1) |
| 69 |
+ (f64::from_bits(ll2i3.1) + f64::from_bits(ll3i4.1)); |
| 70 |
let dl2 = f64::from_bits(ll0i1.2) |
| 71 |
+ f64::from_bits(ll1i2.2) |
| 72 |
+ (f64::from_bits(ll2i3.2) + f64::from_bits(ll3i4.2)); |
| 73 |
dl0 += el0; |
| 74 |
let t12 = f64::from_bits(ACOSH_SINH_REFINE_T1[i1 as usize]) |
| 75 |
* f64::from_bits(ACOSH_ASINH_REFINE_T2[i2 as usize]); |
| 76 |
let t34 = f64::from_bits(ACOSH_SINH_REFINE_T3[i3 as usize]) |
| 77 |
* f64::from_bits(ACOSH_ASINH_REFINE_T4[i4 as usize]); |
| 78 |
let th = t12 * t34; |
| 79 |
let tl = dd_fmla(t12, t34, -th); |
| 80 |
let dh = th * f64::from_bits(t); |
| 81 |
let dl = dd_fmla(th, f64::from_bits(t), -dh); |
| 82 |
let sh = tl * f64::from_bits(t); |
| 83 |
let sl = dd_fmla(tl, f64::from_bits(t), -sh); |
| 84 |
|
| 85 |
let mut dx = DoubleDouble::from_exact_add(dh - 1., dl); |
| 86 |
if z.lo != 0.0 { |
| 87 |
t = z.lo.to_bits(); |
| 88 |
t = t.wrapping_sub((e as i64).wrapping_shl(52) as u64); |
| 89 |
dx.lo += th * f64::from_bits(t); |
| 90 |
} |
| 91 |
dx = DoubleDouble::add(dx, DoubleDouble::new(sl, sh)); |
| 92 |
const CL: [u64; 3] = [0xbfc0000000000000, 0x3fb9999999a0754f, 0xbfb55555555c3157]; |
| 93 |
|
| 94 |
let sl0 = f_fmla(dx.hi, f64::from_bits(CL[2]), f64::from_bits(CL[1])); |
| 95 |
let sl1 = f_fmla(dx.hi, sl0, f64::from_bits(CL[0])); |
| 96 |
|
| 97 |
let sl = dx.hi * sl1; |
| 98 |
const CH: [(u64, u64); 3] = [ |
| 99 |
(0x39024b67ee516e3b, 0x3fe0000000000000), |
| 100 |
(0xb91932ce43199a8d, 0xbfd0000000000000), |
| 101 |
(0x3c655540c15cf91f, 0x3fc5555555555555), |
| 102 |
]; |
| 103 |
let mut s = lpoly_xd_generic(dx, CH, sl); |
| 104 |
s = DoubleDouble::quick_mult(dx, s); |
| 105 |
s = DoubleDouble::add(s, DoubleDouble::new(el2, el1)); |
| 106 |
s = DoubleDouble::add(s, DoubleDouble::new(dl2, dl1)); |
| 107 |
let mut v02 = DoubleDouble::from_exact_add(dl0, s.hi); |
| 108 |
let mut v12 = DoubleDouble::from_exact_add(v02.lo, s.lo); |
| 109 |
let scale = f64::copysign(2., x); |
| 110 |
v02.hi *= scale; |
| 111 |
v12.hi *= scale; |
| 112 |
v12.lo *= scale; |
| 113 |
t = v12.hi.to_bits(); |
| 114 |
if (t & 0x000fffffffffffff) == 0 { |
| 115 |
let w = v12.lo.to_bits(); |
| 116 |
if ((w ^ t) >> 63) != 0 { |
| 117 |
t = t.wrapping_sub(1); |
| 118 |
} else { |
| 119 |
t = t.wrapping_add(1); |
| 120 |
} |
| 121 |
v12.hi = f64::from_bits(t); |
| 122 |
} |
| 123 |
v02.hi + v12.hi |
| 124 |
} |
| 125 |
|
| 126 |
#[cold] |
| 127 |
fn as_asinh_zero(x: f64, x2h: f64, x2l: f64) -> f64 { |
| 128 |
static CH: [(u64, u64); 12] = [ |
| 129 |
(0xbc65555555555555, 0xbfc5555555555555), |
| 130 |
(0x3c499999999949df, 0x3fb3333333333333), |
| 131 |
(0x3c32492496091b0c, 0xbfa6db6db6db6db7), |
| 132 |
(0x3c1c71a35cfa0671, 0x3f9f1c71c71c71c7), |
| 133 |
(0x3c317f937248cf81, 0xbf96e8ba2e8ba2e9), |
| 134 |
(0xbc374e3c1dfd4c3d, 0x3f91c4ec4ec4ec4f), |
| 135 |
(0xbc238e7a467ecc55, 0xbf8c999999999977), |
| 136 |
(0x3c2a83c7bace55eb, 0x3f87a87878786c7e), |
| 137 |
(0xbc2d024df7fa0542, 0xbf83fde50d764083), |
| 138 |
(0xbc2ba9c13deb261f, 0x3f812ef3ceae4d12), |
| 139 |
(0xbc1546da9bc5b32a, 0xbf7df3bd104aa267), |
| 140 |
(0x3c140d284a1d67f9, 0x3f7a685fc5de7a04), |
| 141 |
]; |
| 142 |
|
| 143 |
const CL: [u64; 5] = [ |
| 144 |
0xbf77828d553ec800, |
| 145 |
0x3f751712f7bee368, |
| 146 |
0xbf72e6d98527bcc6, |
| 147 |
0x3f70095da47b392c, |
| 148 |
0xbf63b92d6368192c, |
| 149 |
]; |
| 150 |
|
| 151 |
let yw0 = f_fmla(x2h, f64::from_bits(CL[4]), f64::from_bits(CL[3])); |
| 152 |
let yw1 = f_fmla(x2h, yw0, f64::from_bits(CL[2])); |
| 153 |
let yw2 = f_fmla(x2h, yw1, f64::from_bits(CL[1])); |
| 154 |
|
| 155 |
let y2 = x2h * f_fmla(x2h, yw2, f64::from_bits(CL[0])); |
| 156 |
let mut y1 = lpoly_xd_generic(DoubleDouble::new(x2l, x2h), CH, y2); |
| 157 |
y1 = DoubleDouble::quick_mult(y1, DoubleDouble::new(x2l, x2h)); |
| 158 |
y1 = DoubleDouble::quick_mult_f64(y1, x); |
| 159 |
|
| 160 |
let y0 = DoubleDouble::from_exact_add(x, y1.hi); |
| 161 |
let mut p = DoubleDouble::from_exact_add(y0.lo, y1.lo); |
| 162 |
|
| 163 |
let mut t = p.hi.to_bits(); |
| 164 |
if (t & 0x000fffffffffffff) == 0 { |
| 165 |
let w = p.lo.to_bits(); |
| 166 |
if ((w ^ t) >> 63) != 0 { |
| 167 |
t = t.wrapping_sub(1); |
| 168 |
} else { |
| 169 |
t = t.wrapping_add(1); |
| 170 |
} |
| 171 |
p.hi = f64::from_bits(t); |
| 172 |
} |
| 173 |
y0.hi + p.hi |
| 174 |
} |
| 175 |
|
| 176 |
/// Huperbolic sine function |
| 177 |
/// |
| 178 |
/// Max ULP 0.5 |
| 179 |
pub fn f_asinh(x: f64) -> f64 { |
| 180 |
let ax = x.abs(); |
| 181 |
let ix = ax.to_bits(); |
| 182 |
let u = ix; |
| 183 |
if u < 0x3fbb000000000000u64 { |
| 184 |
// |x| < 0x1.bp-4 |
| 185 |
// for |x| < 0x1.7137449123ef7p-26, asinh(x) rounds to x to nearest |
| 186 |
// for |x| < 0x1p-1022 we have underflow but not for 0x1p-1022 (to nearest) |
| 187 |
if u < 0x3e57137449123ef7u64 { |
| 188 |
// |x| < 0x1.7137449123ef7p-26 |
| 189 |
if u == 0 { |
| 190 |
return x; |
| 191 |
} |
| 192 |
let res = f_fmla(f64::from_bits(0xbc30000000000000), x, x); |
| 193 |
return res; |
| 194 |
} |
| 195 |
let x2h = x * x; |
| 196 |
let x2l = dd_fmla(x, x, -x2h); |
| 197 |
let x3h = x2h * x; |
| 198 |
let sl; |
| 199 |
if u < 0x3f93000000000000u64 { |
| 200 |
// |x| < 0x1.3p-6 |
| 201 |
if u < 0x3f30000000000000u64 { |
| 202 |
// |x| < 0x1p-12 |
| 203 |
if u < 0x3e5a000000000000u64 { |
| 204 |
// |x| < 0x1.ap-26 |
| 205 |
sl = x3h * f64::from_bits(0xbfc5555555555555); |
| 206 |
} else { |
| 207 |
const CL: [u64; 2] = [0xbfc5555555555555, 0x3fb3333327c57c60]; |
| 208 |
let sl0 = f_fmla(x2h, f64::from_bits(CL[1]), f64::from_bits(CL[0])); |
| 209 |
sl = x3h * sl0; |
| 210 |
} |
| 211 |
} else { |
| 212 |
const CL: [u64; 4] = [ |
| 213 |
0xbfc5555555555555, |
| 214 |
0x3fb333333332f2ff, |
| 215 |
0xbfa6db6d9a665159, |
| 216 |
0x3f9f186866d775f0, |
| 217 |
]; |
| 218 |
let sl0 = f_fmla(x2h, f64::from_bits(CL[3]), f64::from_bits(CL[2])); |
| 219 |
let sl1 = f_fmla(x2h, sl0, f64::from_bits(CL[1])); |
| 220 |
sl = x3h * f_fmla(x2h, sl1, f64::from_bits(CL[0])); |
| 221 |
} |
| 222 |
} else { |
| 223 |
const CL: [u64; 7] = [ |
| 224 |
0xbfc5555555555555, |
| 225 |
0x3fb3333333333310, |
| 226 |
0xbfa6db6db6da466c, |
| 227 |
0x3f9f1c71c2ea7be4, |
| 228 |
0xbf96e8b651b09d72, |
| 229 |
0x3f91c309fc0e69c2, |
| 230 |
0xbf8bab7833c1e000, |
| 231 |
]; |
| 232 |
let c1 = f_fmla(x2h, f64::from_bits(CL[2]), f64::from_bits(CL[1])); |
| 233 |
let c3 = f_fmla(x2h, f64::from_bits(CL[4]), f64::from_bits(CL[3])); |
| 234 |
let c5 = f_fmla(x2h, f64::from_bits(CL[6]), f64::from_bits(CL[5])); |
| 235 |
let x4 = x2h * x2h; |
| 236 |
|
| 237 |
let sl0 = f_fmla(x4, c5, c3); |
| 238 |
let sl1 = f_fmla(x4, sl0, c1); |
| 239 |
|
| 240 |
sl = x3h * f_fmla(x2h, sl1, f64::from_bits(CL[0])); |
| 241 |
} |
| 242 |
let eps = f64::from_bits(0x3ca6000000000000) * x3h; |
| 243 |
let lb = x + (sl - eps); |
| 244 |
let ub = x + (sl + eps); |
| 245 |
if lb == ub { |
| 246 |
return lb; |
| 247 |
} |
| 248 |
return as_asinh_zero(x, x2h, x2l); |
| 249 |
} |
| 250 |
|
| 251 |
// |x| >= 0x1.bp-4 |
| 252 |
let mut x2h: f64 = 0.; |
| 253 |
let mut x2l: f64 = 0.; |
| 254 |
let mut off: i32 = 0x3ff; |
| 255 |
|
| 256 |
let va: DoubleDouble = if u < 0x4190000000000000 { |
| 257 |
// x < 0x1p+26 |
| 258 |
x2h = x * x; |
| 259 |
x2l = dd_fmla(x, x, -x2h); |
| 260 |
let mut dt = if u < 0x3ff0000000000000 { |
| 261 |
DoubleDouble::from_exact_add(1., x2h) |
| 262 |
} else { |
| 263 |
DoubleDouble::from_exact_add(x2h, 1.) |
| 264 |
}; |
| 265 |
dt.lo += x2l; |
| 266 |
|
| 267 |
let ah = dt.hi.sqrt(); |
| 268 |
let rs = 0.5 / dt.hi; |
| 269 |
let al = (dt.lo - dd_fmla(ah, ah, -dt.hi)) * (rs * ah); |
| 270 |
let mut ma = DoubleDouble::from_exact_add(ah, ax); |
| 271 |
ma.lo += al; |
| 272 |
ma |
| 273 |
} else if u < 0x4330000000000000 { |
| 274 |
DoubleDouble::new(0.5 / ax, 2. * ax) |
| 275 |
} else { |
| 276 |
if u >= 0x7ff0000000000000u64 { |
| 277 |
return x + x; |
| 278 |
} // +-inf or nan |
| 279 |
off = 0x3fe; |
| 280 |
DoubleDouble::new(0., ax) |
| 281 |
}; |
| 282 |
|
| 283 |
let mut t = va.hi.to_bits(); |
| 284 |
let ex = (t >> 52) as i32; |
| 285 |
let e = ex.wrapping_sub(off); |
| 286 |
t &= 0x000fffffffffffff; |
| 287 |
let ed = e as f64; |
| 288 |
let i = t >> (52 - 5); |
| 289 |
let d = (t & 0x00007fffffffffff) as i64; |
| 290 |
let b_i = ACOSH_ASINH_B[i as usize]; |
| 291 |
let j: u64 = t |
| 292 |
.wrapping_add((b_i[0] as u64).wrapping_shl(33)) |
| 293 |
.wrapping_add((b_i[1] as i64).wrapping_mul(d >> 16) as u64) |
| 294 |
>> (52 - 10); |
| 295 |
t |= 0x3ffu64 << 52; |
| 296 |
let i1 = (j >> 5) as i32; |
| 297 |
let i2 = j & 0x1f; |
| 298 |
let r = |
| 299 |
f64::from_bits(ACOSH_ASINH_R1[i1 as usize]) * f64::from_bits(ACOSH_ASINH_R2[i2 as usize]); |
| 300 |
let dx = dd_fmla(r, f64::from_bits(t), -1.); |
| 301 |
let dx2 = dx * dx; |
| 302 |
const C: [u64; 5] = [ |
| 303 |
0xbfe0000000000000, |
| 304 |
0x3fd5555555555530, |
| 305 |
0xbfcfffffffffffa0, |
| 306 |
0x3fc99999e33a6366, |
| 307 |
0xbfc555559ef9525f, |
| 308 |
]; |
| 309 |
|
| 310 |
let fw0 = f_fmla(dx, f64::from_bits(C[3]), f64::from_bits(C[2])); |
| 311 |
let fw1 = f_fmla(dx, f64::from_bits(C[1]), f64::from_bits(C[0])); |
| 312 |
let fw2 = f_fmla(dx2, f64::from_bits(C[4]), fw0); |
| 313 |
|
| 314 |
let f = dx2 * f_fmla(dx2, fw2, fw1); |
| 315 |
const L2H: f64 = f64::from_bits(0x3fe62e42fefa3800); |
| 316 |
const L2L: f64 = f64::from_bits(0x3d2ef35793c76730); |
| 317 |
let l1i1 = ACOSH_ASINH_L1[i1 as usize]; |
| 318 |
let l1i2 = ACOSH_ASINH_L2[i2 as usize]; |
| 319 |
let mut lh = f_fmla(L2H, ed, f64::from_bits(l1i1.1) + f64::from_bits(l1i2.1)); |
| 320 |
let mut ll = f_fmla( |
| 321 |
L2L, |
| 322 |
ed, |
| 323 |
f64::from_bits(l1i1.0) + f64::from_bits(l1i2.0) + va.lo / va.hi + f, |
| 324 |
); |
| 325 |
ll += dx; |
| 326 |
lh *= f64::copysign(1., x); |
| 327 |
ll *= f64::copysign(1., x); |
| 328 |
let eps = 1.63e-19; |
| 329 |
let lb = lh + (ll - eps); |
| 330 |
let ub = lh + (ll + eps); |
| 331 |
if lb == ub { |
| 332 |
return lb; |
| 333 |
} |
| 334 |
if ax < f64::from_bits(0x3fd0000000000000) { |
| 335 |
return as_asinh_zero(x, x2h, x2l); |
| 336 |
} |
| 337 |
asinh_refine(x, f64::from_bits(0x3ff71547652b82fe) * lb.abs(), va) |
| 338 |
} |
| 339 |
|
| 340 |
#[cfg(test)] |
| 341 |
mod tests { |
| 342 |
use super::*; |
| 343 |
|
| 344 |
#[test] |
| 345 |
fn test_asinh() { |
| 346 |
assert_eq!(f_asinh(-0.05268859863273256), -0.05266425100170862); |
| 347 |
assert_eq!(f_asinh(1.05268859863273256), 0.9181436936151385); |
| 348 |
} |
| 349 |
} |