| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 8/2025. 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::{f_fmla, is_integer}; |
| 30 |
use crate::double_double::DoubleDouble; |
| 31 |
use crate::logs::fast_log_dd; |
| 32 |
use crate::polyeval::{f_polyeval6, f_polyeval7, f_polyeval8}; |
| 33 |
use crate::pow_exec::exp_dd_fast; |
| 34 |
use crate::rounding::CpuFloor; |
| 35 |
use crate::sincospi::f_fast_sinpi_dd; |
| 36 |
|
| 37 |
/// Computes gamma(x) |
| 38 |
/// |
| 39 |
/// ulp 1 |
| 40 |
pub fn f_tgamma(x: f64) -> f64 { |
| 41 |
let x_a = f64::from_bits(x.to_bits() & 0x7fff_ffff_ffff_ffff); |
| 42 |
|
| 43 |
if !x.is_normal() { |
| 44 |
if x == 0.0 { |
| 45 |
return f64::INFINITY; |
| 46 |
} |
| 47 |
|
| 48 |
if x.is_nan() { |
| 49 |
return x + x; |
| 50 |
} |
| 51 |
|
| 52 |
if x.is_infinite() { |
| 53 |
if x.is_sign_negative() { |
| 54 |
return f64::NAN; |
| 55 |
} |
| 56 |
return x; |
| 57 |
} |
| 58 |
} |
| 59 |
|
| 60 |
if x >= 171.624 { |
| 61 |
return f64::INFINITY; |
| 62 |
} |
| 63 |
|
| 64 |
if is_integer(x) { |
| 65 |
if x < 0. { |
| 66 |
return f64::NAN; |
| 67 |
} |
| 68 |
if x < 38. { |
| 69 |
let mut t = DoubleDouble::new(0., 1.); |
| 70 |
let k = x as i64; |
| 71 |
let mut x0 = 1.0; |
| 72 |
for _i in 1..k { |
| 73 |
t = DoubleDouble::quick_mult_f64(t, x0); |
| 74 |
t = DoubleDouble::from_exact_add(t.hi, t.lo); |
| 75 |
x0 += 1.0; |
| 76 |
} |
| 77 |
return t.to_f64(); |
| 78 |
} |
| 79 |
} |
| 80 |
|
| 81 |
if x <= -184.0 { |
| 82 |
/* negative non-integer */ |
| 83 |
/* For x <= -184, x non-integer, |gamma(x)| < 2^-1078. */ |
| 84 |
static SIGN: [f64; 2] = [ |
| 85 |
f64::from_bits(0x0010000000000000), |
| 86 |
f64::from_bits(0x8010000000000000), |
| 87 |
]; |
| 88 |
let k = x as i64; |
| 89 |
return f64::from_bits(0x0010000000000000) * SIGN[((k & 1) != 0) as usize]; |
| 90 |
} |
| 91 |
|
| 92 |
const EULER_DD: DoubleDouble = |
| 93 |
DoubleDouble::from_bit_pair((0xbc56cb90701fbfab, 0x3fe2788cfc6fb619)); |
| 94 |
|
| 95 |
if x_a < 0.006 { |
| 96 |
if x_a.to_bits() < (0x71e0000000000000u64 >> 1) { |
| 97 |
// |x| < 0x1p-112 |
| 98 |
return 1. / x; |
| 99 |
} |
| 100 |
if x_a < 2e-10 { |
| 101 |
// x is tiny then Gamma(x) = 1/x - euler |
| 102 |
let p = DoubleDouble::full_dd_sub(DoubleDouble::from_quick_recip(x), EULER_DD); |
| 103 |
return p.to_f64(); |
| 104 |
} else if x_a < 2e-6 { |
| 105 |
// x is small then Gamma(x) = 1/x - euler + a2*x |
| 106 |
// a2 = 1/12 * (6 * euler^2 + pi^2) |
| 107 |
const A2: DoubleDouble = |
| 108 |
DoubleDouble::from_bit_pair((0x3c8dd92b465a8221, 0x3fefa658c23b1578)); |
| 109 |
let rcp = DoubleDouble::from_quick_recip(x); |
| 110 |
let p = DoubleDouble::full_dd_add(DoubleDouble::mul_f64_add(A2, x, -EULER_DD), rcp); |
| 111 |
return p.to_f64(); |
| 112 |
} |
| 113 |
|
| 114 |
// Laurent series of Gamma(x) |
| 115 |
const C: [(u64, u64); 8] = [ |
| 116 |
(0x3c8dd92b465a8221, 0x3fefa658c23b1578), |
| 117 |
(0x3c53a4f483760950, 0xbfed0a118f324b63), |
| 118 |
(0x3c7fabe4f7369157, 0x3fef6a51055096b5), |
| 119 |
(0x3c8c9fc795fc6142, 0xbfef6c80ec38b67b), |
| 120 |
(0xbc5042339d62e721, 0x3fefc7e0a6eb310b), |
| 121 |
(0xbc86fd0d8bdc0c1e, 0xbfefdf3f157b7a39), |
| 122 |
(0xbc89b912df09395d, 0x3feff07b5a17ff6c), |
| 123 |
(0x3c4e626faf780ff9, 0xbfeff803d68a0bd4), |
| 124 |
]; |
| 125 |
let rcp = DoubleDouble::from_quick_recip(x); |
| 126 |
let mut p = DoubleDouble::mul_f64_add( |
| 127 |
DoubleDouble::from_bit_pair(C[7]), |
| 128 |
x, |
| 129 |
DoubleDouble::from_bit_pair(C[6]), |
| 130 |
); |
| 131 |
p = DoubleDouble::mul_f64_add(p, x, DoubleDouble::from_bit_pair(C[5])); |
| 132 |
p = DoubleDouble::mul_f64_add(p, x, DoubleDouble::from_bit_pair(C[4])); |
| 133 |
p = DoubleDouble::mul_f64_add(p, x, DoubleDouble::from_bit_pair(C[3])); |
| 134 |
p = DoubleDouble::mul_f64_add(p, x, DoubleDouble::from_bit_pair(C[2])); |
| 135 |
p = DoubleDouble::mul_f64_add(p, x, DoubleDouble::from_bit_pair(C[1])); |
| 136 |
p = DoubleDouble::mul_f64_add(p, x, DoubleDouble::from_bit_pair(C[0])); |
| 137 |
let z = DoubleDouble::mul_f64_add(p, x, DoubleDouble::full_dd_sub(rcp, EULER_DD)); |
| 138 |
return z.to_f64(); |
| 139 |
} |
| 140 |
|
| 141 |
let mut fact = DoubleDouble::new(0., 0.0f64); |
| 142 |
let mut parity = 1.0; |
| 143 |
const PI: DoubleDouble = DoubleDouble::from_bit_pair((0x3ca1a62633145c07, 0x400921fb54442d18)); |
| 144 |
let mut dy = DoubleDouble::new(0., x); |
| 145 |
let mut result: DoubleDouble; |
| 146 |
|
| 147 |
// reflection |
| 148 |
if dy.hi < 0. { |
| 149 |
if dy.hi.cpu_floor() == dy.hi { |
| 150 |
return f64::NAN; |
| 151 |
} |
| 152 |
dy.hi = f64::from_bits(dy.hi.to_bits() & 0x7fff_ffff_ffff_ffff); |
| 153 |
let y1 = x_a.cpu_floor(); |
| 154 |
let fraction = x_a - y1; |
| 155 |
if fraction != 0.0 |
| 156 |
// is it an integer? |
| 157 |
{ |
| 158 |
// is it odd or even? |
| 159 |
if y1 != (y1 * 0.5).cpu_floor() * 2.0 { |
| 160 |
parity = -1.0; |
| 161 |
} |
| 162 |
fact = DoubleDouble::div(-PI, f_fast_sinpi_dd(fraction)); |
| 163 |
fact = DoubleDouble::from_exact_add(fact.hi, fact.lo); |
| 164 |
dy = DoubleDouble::from_full_exact_add(dy.hi, 1.0); |
| 165 |
} |
| 166 |
} |
| 167 |
|
| 168 |
if dy.hi < 12.0 { |
| 169 |
let y1 = dy; |
| 170 |
let z: DoubleDouble; |
| 171 |
let mut n = 0; |
| 172 |
// x is in (0.06, 1.0). |
| 173 |
if dy.hi < 1.0 { |
| 174 |
z = dy; |
| 175 |
dy = DoubleDouble::full_add_f64(dy, 1.0); |
| 176 |
} else |
| 177 |
// x is in [1.0, max]. |
| 178 |
{ |
| 179 |
n = dy.hi as i32 - 1; |
| 180 |
dy = DoubleDouble::full_add_f64(dy, -n as f64); |
| 181 |
z = DoubleDouble::full_add_f64(dy, -1.0); |
| 182 |
} |
| 183 |
|
| 184 |
// Gamma(x+1) on [1;2] generated by Wolfram Mathematica: |
| 185 |
// <<FunctionApproximations` |
| 186 |
// ClearAll["Global`*"] |
| 187 |
// f[x_]:=Gamma[x+1] |
| 188 |
// {err0, approx}=MiniMaxApproximation[f[z],{z,{0,1 },9,8},WorkingPrecision->90] |
| 189 |
// num=Numerator[approx][[1]]; |
| 190 |
// den=Denominator[approx][[1]]; |
| 191 |
// poly=den; |
| 192 |
// coeffs=CoefficientList[poly,z]; |
| 193 |
// TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50}, ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] |
| 194 |
let ps_num = f_polyeval8( |
| 195 |
z.hi, |
| 196 |
f64::from_bits(0x3fb38b80568a42aa), |
| 197 |
f64::from_bits(0xbf8e7685b00d63a6), |
| 198 |
f64::from_bits(0xbf80629ed2c48f1a), |
| 199 |
f64::from_bits(0xbf6dfc4cdbcee96a), |
| 200 |
f64::from_bits(0xbf471816939dc42b), |
| 201 |
f64::from_bits(0xbf24bede7d8b3c20), |
| 202 |
f64::from_bits(0xbeef56936d891e42), |
| 203 |
f64::from_bits(0xbec3e2b405350813), |
| 204 |
); |
| 205 |
let mut p_num = DoubleDouble::mul_f64_add( |
| 206 |
z, |
| 207 |
ps_num, |
| 208 |
DoubleDouble::from_bit_pair((0xbc700f441aea2edb, 0x3fd218bdde7878b8)), |
| 209 |
); |
| 210 |
p_num = DoubleDouble::mul_add( |
| 211 |
z, |
| 212 |
p_num, |
| 213 |
DoubleDouble::from_bit_pair((0x3b7056a45a2fa50e, 0x3ff0000000000000)), |
| 214 |
); |
| 215 |
p_num = DoubleDouble::from_exact_add(p_num.hi, p_num.lo); |
| 216 |
|
| 217 |
let ps_den = f_polyeval7( |
| 218 |
z.hi, |
| 219 |
f64::from_bits(0xbfdaa4f09f0caab1), |
| 220 |
f64::from_bits(0xbfc960ba48423f9d), |
| 221 |
f64::from_bits(0x3fb6873b64e8ccd6), |
| 222 |
f64::from_bits(0x3f69ea1ca5b8a225), |
| 223 |
f64::from_bits(0xbf77b166f68a2e63), |
| 224 |
f64::from_bits(0x3f4fd1eff9193728), |
| 225 |
f64::from_bits(0xbf0c1a43f4985c97), |
| 226 |
); |
| 227 |
let mut p_den = DoubleDouble::mul_f64_add( |
| 228 |
z, |
| 229 |
ps_den, |
| 230 |
DoubleDouble::from_bit_pair((0xbc759594c51ad8b7, 0x3feb84ebebabf275)), |
| 231 |
); |
| 232 |
p_den = DoubleDouble::mul_add_f64(z, p_den, f64::from_bits(0x3ff0000000000000)); |
| 233 |
p_den = DoubleDouble::from_exact_add(p_den.hi, p_den.lo); |
| 234 |
result = DoubleDouble::div(p_num, p_den); |
| 235 |
if y1.hi < dy.hi { |
| 236 |
result = DoubleDouble::div(result, y1); |
| 237 |
} else if y1.hi > dy.hi { |
| 238 |
for _ in 0..n { |
| 239 |
result = DoubleDouble::mult(result, dy); |
| 240 |
dy = DoubleDouble::full_add_f64(dy, 1.0); |
| 241 |
} |
| 242 |
} |
| 243 |
} else { |
| 244 |
if x > 171.624e+0 { |
| 245 |
return f64::INFINITY; |
| 246 |
} |
| 247 |
// Stirling's approximation of Log(Gamma) and then Exp[Log[Gamma]] |
| 248 |
let y_recip = dy.recip(); |
| 249 |
let y_sqr = DoubleDouble::mult(y_recip, y_recip); |
| 250 |
// Bernoulli coefficients generated by SageMath: |
| 251 |
// var('x') |
| 252 |
// def bernoulli_terms(x, N): |
| 253 |
// S = 0 |
| 254 |
// for k in range(1, N+1): |
| 255 |
// B = bernoulli(2*k) |
| 256 |
// term = (B / (2*k*(2*k-1))) * x**((2*k-1)) |
| 257 |
// S += term |
| 258 |
// return S |
| 259 |
// |
| 260 |
// terms = bernoulli_terms(x, 7) |
| 261 |
let bernoulli_poly_s = f_polyeval6( |
| 262 |
y_sqr.hi, |
| 263 |
f64::from_bits(0xbf66c16c16c16c17), |
| 264 |
f64::from_bits(0x3f4a01a01a01a01a), |
| 265 |
f64::from_bits(0xbf43813813813814), |
| 266 |
f64::from_bits(0x3f4b951e2b18ff23), |
| 267 |
f64::from_bits(0xbf5f6ab0d9993c7d), |
| 268 |
f64::from_bits(0x3f7a41a41a41a41a), |
| 269 |
); |
| 270 |
let bernoulli_poly = DoubleDouble::mul_f64_add( |
| 271 |
y_sqr, |
| 272 |
bernoulli_poly_s, |
| 273 |
DoubleDouble::from_bit_pair((0x3c55555555555555, 0x3fb5555555555555)), |
| 274 |
); |
| 275 |
// Log[Gamma(x)] = x*log(x) - x + 1/2*Log(2*PI/x) + bernoulli_terms |
| 276 |
const LOG2_PI_OVER_2: DoubleDouble = |
| 277 |
DoubleDouble::from_bit_pair((0xbc865b5a1b7ff5df, 0x3fed67f1c864beb5)); |
| 278 |
let mut log_gamma = DoubleDouble::add( |
| 279 |
DoubleDouble::mul_add(bernoulli_poly, y_recip, -dy), |
| 280 |
LOG2_PI_OVER_2, |
| 281 |
); |
| 282 |
let dy_log = fast_log_dd(dy); |
| 283 |
log_gamma = DoubleDouble::mul_add( |
| 284 |
DoubleDouble::from_exact_add(dy_log.hi, dy_log.lo), |
| 285 |
DoubleDouble::add_f64(dy, -0.5), |
| 286 |
log_gamma, |
| 287 |
); |
| 288 |
let log_prod = log_gamma.to_f64(); |
| 289 |
if log_prod >= 690. { |
| 290 |
// underflow/overflow case |
| 291 |
log_gamma = DoubleDouble::quick_mult_f64(log_gamma, 0.5); |
| 292 |
result = exp_dd_fast(log_gamma); |
| 293 |
let exp_result = result; |
| 294 |
result.hi *= parity; |
| 295 |
result.lo *= parity; |
| 296 |
if fact.lo != 0. && fact.hi != 0. { |
| 297 |
// y / x = y / (z*z) = y / z * 1/z |
| 298 |
result = DoubleDouble::from_exact_add(result.hi, result.lo); |
| 299 |
result = DoubleDouble::div(fact, result); |
| 300 |
result = DoubleDouble::div(result, exp_result); |
| 301 |
} else { |
| 302 |
result = DoubleDouble::quick_mult(result, exp_result); |
| 303 |
} |
| 304 |
|
| 305 |
let err = f_fmla( |
| 306 |
result.hi.abs(), |
| 307 |
f64::from_bits(0x3c20000000000000), // 2^-61 |
| 308 |
f64::from_bits(0x3bd0000000000000), // 2^-65 |
| 309 |
); |
| 310 |
let ub = result.hi + (result.lo + err); |
| 311 |
let lb = result.hi + (result.lo - err); |
| 312 |
if ub == lb { |
| 313 |
return result.to_f64(); |
| 314 |
} |
| 315 |
return result.to_f64(); |
| 316 |
} |
| 317 |
result = exp_dd_fast(log_gamma); |
| 318 |
} |
| 319 |
|
| 320 |
if fact.lo != 0. && fact.hi != 0. { |
| 321 |
// y / x = y / (z*z) = y / z * 1/z |
| 322 |
result = DoubleDouble::from_exact_add(result.hi, result.lo); |
| 323 |
result = DoubleDouble::div(fact, result); |
| 324 |
} |
| 325 |
result.to_f64() * parity |
| 326 |
} |
| 327 |
|
| 328 |
#[cfg(test)] |
| 329 |
mod tests { |
| 330 |
use super::*; |
| 331 |
|
| 332 |
#[test] |
| 333 |
fn test_tgamma() { |
| 334 |
// assert_eq!(f_tgamma(6.757812502211891), 459.54924419209556); |
| 335 |
assert_eq!(f_tgamma(-1.70000000042915), 2.513923520668069); |
| 336 |
assert_eq!(f_tgamma(5.), 24.); |
| 337 |
assert_eq!(f_tgamma(24.), 25852016738884980000000.); |
| 338 |
assert_eq!(f_tgamma(6.4324324), 255.1369211339094); |
| 339 |
assert_eq!(f_tgamma(f64::INFINITY), f64::INFINITY); |
| 340 |
assert_eq!(f_tgamma(0.), f64::INFINITY); |
| 341 |
assert_eq!(f_tgamma(-0.), f64::INFINITY); |
| 342 |
assert!(f_tgamma(f64::NAN).is_nan()); |
| 343 |
} |
| 344 |
} |