| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 6/2025. 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::acospi::INV_PI_DD; |
| 30 |
use crate::asin::asin_eval; |
| 31 |
use crate::asin_eval_dyadic::asin_eval_dyadic; |
| 32 |
use crate::common::{dd_fmla, dyad_fmla, f_fmla}; |
| 33 |
use crate::double_double::DoubleDouble; |
| 34 |
use crate::dyadic_float::{DyadicFloat128, DyadicSign}; |
| 35 |
use crate::rounding::CpuRound; |
| 36 |
|
| 37 |
/// Computes asin(x)/PI |
| 38 |
/// |
| 39 |
/// Max found ULP 0.5 |
| 40 |
pub fn f_asinpi(x: f64) -> f64 { |
| 41 |
let x_e = (x.to_bits() >> 52) & 0x7ff; |
| 42 |
const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; |
| 43 |
|
| 44 |
let x_abs = f64::from_bits(x.to_bits() & 0x7fff_ffff_ffff_ffff); |
| 45 |
|
| 46 |
// |x| < 0.5. |
| 47 |
if x_e < E_BIAS - 1 { |
| 48 |
// |x| < 2^-26. |
| 49 |
if x_e < E_BIAS - 26 { |
| 50 |
// When |x| < 2^-26, the relative error of the approximation asin(x) ~ x |
| 51 |
// is: |
| 52 |
// |asin(x) - x| / |asin(x)| < |x^3| / (6|x|) |
| 53 |
// = x^2 / 6 |
| 54 |
// < 2^-54 |
| 55 |
// < epsilon(1)/2. |
| 56 |
// = x otherwise. , |
| 57 |
if x.abs() == 0. { |
| 58 |
return x; |
| 59 |
} |
| 60 |
|
| 61 |
if x_e < E_BIAS - 56 { |
| 62 |
if (x_abs.to_bits().wrapping_shl(12)) == 0x59af9a1194efe000u64 { |
| 63 |
let e = (x.to_bits() >> 52) & 0x7ff; |
| 64 |
let h = f64::from_bits(0x3c7b824198b94a89); |
| 65 |
let l = f64::from_bits(0x391fffffffffffff); |
| 66 |
let mut t = (if x > 0. { 1.0f64 } else { -1.0f64 }).to_bits(); |
| 67 |
t = t.wrapping_sub(0x3c9u64.wrapping_sub(e).wrapping_shl(52)); |
| 68 |
return f_fmla(l, f64::from_bits(t), h * f64::from_bits(t)); |
| 69 |
} |
| 70 |
|
| 71 |
let h = x * INV_PI_DD.hi; |
| 72 |
let sx = x * f64::from_bits(0x4690000000000000); /* scale x */ |
| 73 |
let mut l = dd_fmla(sx, INV_PI_DD.hi, -h * f64::from_bits(0x4690000000000000)); |
| 74 |
l = dd_fmla(sx, INV_PI_DD.lo, l); |
| 75 |
/* scale back */ |
| 76 |
let res = dyad_fmla(l, f64::from_bits(0x3950000000000000), h); |
| 77 |
return res; |
| 78 |
} |
| 79 |
|
| 80 |
/* We use the Sollya polynomial 0x1.45f306dc9c882a53f84eafa3ea4p-2 * x |
| 81 |
+ 0x1.b2995e7b7b606p-5 * x^3, with relative error bounded by 2^-106.965 |
| 82 |
on [2^-53, 2^-26] */ |
| 83 |
const C1H: f64 = f64::from_bits(0x3fd45f306dc9c883); |
| 84 |
const C1L: f64 = f64::from_bits(0xbc76b01ec5417057); |
| 85 |
const C3: f64 = f64::from_bits(0x3fab2995e7b7b606); |
| 86 |
let h = C1H; |
| 87 |
let l = dd_fmla(C3, x * x, C1L); |
| 88 |
/* multiply h+l by x */ |
| 89 |
let hh = h * x; |
| 90 |
let mut ll = dd_fmla(h, x, -hh); |
| 91 |
/* hh+ll = h*x */ |
| 92 |
ll = dd_fmla(l, x, ll); |
| 93 |
return hh + ll; |
| 94 |
} |
| 95 |
|
| 96 |
let x_sq = DoubleDouble::from_exact_mult(x, x); |
| 97 |
let err = x_abs * f64::from_bits(0x3cc0000000000000); |
| 98 |
// Polynomial approximation: |
| 99 |
// p ~ asin(x)/x |
| 100 |
|
| 101 |
let (p, err) = asin_eval(x_sq, err); |
| 102 |
// asin(x) ~ x * (ASIN_COEFFS[idx][0] + p) |
| 103 |
let mut r0 = DoubleDouble::from_exact_mult(x, p.hi); |
| 104 |
let mut r_lo = f_fmla(x, p.lo, r0.lo); |
| 105 |
|
| 106 |
r0 = DoubleDouble::mult(DoubleDouble::new(r_lo, r0.hi), INV_PI_DD); |
| 107 |
r_lo = r0.lo; |
| 108 |
|
| 109 |
let r_upper = r0.hi + (r_lo + err); |
| 110 |
let r_lower = r0.hi + (r_lo - err); |
| 111 |
|
| 112 |
if r_upper == r_lower { |
| 113 |
return r_upper; |
| 114 |
} |
| 115 |
|
| 116 |
// Ziv's accuracy test failed, perform 128-bit calculation. |
| 117 |
|
| 118 |
// Recalculate mod 1/64. |
| 119 |
let idx = (x_sq.hi * f64::from_bits(0x4050000000000000)).cpu_round() as usize; |
| 120 |
|
| 121 |
// Get x^2 - idx/64 exactly. When FMA is available, double-double |
| 122 |
// multiplication will be correct for all rounding modes. Otherwise, we use |
| 123 |
// Float128 directly. |
| 124 |
let x_f128 = DyadicFloat128::new_from_f64(x); |
| 125 |
|
| 126 |
let u: DyadicFloat128; |
| 127 |
#[cfg(any( |
| 128 |
all( |
| 129 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 130 |
target_feature = "fma" |
| 131 |
), |
| 132 |
target_arch = "aarch64" |
| 133 |
))] |
| 134 |
{ |
| 135 |
// u = x^2 - idx/64 |
| 136 |
let u_hi = DyadicFloat128::new_from_f64(f_fmla( |
| 137 |
idx as f64, |
| 138 |
f64::from_bits(0xbf90000000000000), |
| 139 |
x_sq.hi, |
| 140 |
)); |
| 141 |
u = u_hi.quick_add(&DyadicFloat128::new_from_f64(x_sq.lo)); |
| 142 |
} |
| 143 |
|
| 144 |
#[cfg(not(any( |
| 145 |
all( |
| 146 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 147 |
target_feature = "fma" |
| 148 |
), |
| 149 |
target_arch = "aarch64" |
| 150 |
)))] |
| 151 |
{ |
| 152 |
let x_sq_f128 = x_f128.quick_mul(&x_f128); |
| 153 |
u = x_sq_f128.quick_add(&DyadicFloat128::new_from_f64( |
| 154 |
idx as f64 * (f64::from_bits(0xbf90000000000000)), |
| 155 |
)); |
| 156 |
} |
| 157 |
|
| 158 |
let p_f128 = asin_eval_dyadic(u, idx); |
| 159 |
let mut r = x_f128.quick_mul(&p_f128); |
| 160 |
r = r.quick_mul(&crate::acospi::INV_PI_F128); |
| 161 |
return r.fast_as_f64(); |
| 162 |
} |
| 163 |
|
| 164 |
const PI_OVER_TWO: DoubleDouble = DoubleDouble::new( |
| 165 |
f64::from_bits(0x3c91a62633145c07), |
| 166 |
f64::from_bits(0x3ff921fb54442d18), |
| 167 |
); |
| 168 |
|
| 169 |
let x_sign = if x.is_sign_negative() { -1.0 } else { 1.0 }; |
| 170 |
|
| 171 |
// |x| >= 1 |
| 172 |
if x_e >= E_BIAS { |
| 173 |
// x = +-1, asin(x) = +- pi/2 |
| 174 |
if x_abs == 1.0 { |
| 175 |
// return +- pi/2 |
| 176 |
return x * 0.5; // asinpi_specific |
| 177 |
} |
| 178 |
// |x| > 1, return NaN. |
| 179 |
if x.is_nan() { |
| 180 |
return x; |
| 181 |
} |
| 182 |
return f64::NAN; |
| 183 |
} |
| 184 |
|
| 185 |
// u = (1 - |x|)/2 |
| 186 |
let u = f_fmla(x_abs, -0.5, 0.5); |
| 187 |
// v_hi + v_lo ~ sqrt(u). |
| 188 |
// Let: |
| 189 |
// h = u - v_hi^2 = (sqrt(u) - v_hi) * (sqrt(u) + v_hi) |
| 190 |
// Then: |
| 191 |
// sqrt(u) = v_hi + h / (sqrt(u) + v_hi) |
| 192 |
// ~ v_hi + h / (2 * v_hi) |
| 193 |
// So we can use: |
| 194 |
// v_lo = h / (2 * v_hi). |
| 195 |
// Then, |
| 196 |
// asin(x) ~ pi/2 - 2*(v_hi + v_lo) * P(u) |
| 197 |
let v_hi = u.sqrt(); |
| 198 |
let h; |
| 199 |
#[cfg(any( |
| 200 |
all( |
| 201 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 202 |
target_feature = "fma" |
| 203 |
), |
| 204 |
target_arch = "aarch64" |
| 205 |
))] |
| 206 |
{ |
| 207 |
h = f_fmla(v_hi, -v_hi, u); |
| 208 |
} |
| 209 |
#[cfg(not(any( |
| 210 |
all( |
| 211 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 212 |
target_feature = "fma" |
| 213 |
), |
| 214 |
target_arch = "aarch64" |
| 215 |
)))] |
| 216 |
{ |
| 217 |
let v_hi_sq = DoubleDouble::from_exact_mult(v_hi, v_hi); |
| 218 |
h = (u - v_hi_sq.hi) - v_hi_sq.lo; |
| 219 |
} |
| 220 |
// Scale v_lo and v_hi by 2 from the formula: |
| 221 |
// vh = v_hi * 2 |
| 222 |
// vl = 2*v_lo = h / v_hi. |
| 223 |
let vh = v_hi * 2.0; |
| 224 |
let vl = h / v_hi; |
| 225 |
|
| 226 |
// Polynomial approximation: |
| 227 |
// p ~ asin(sqrt(u))/sqrt(u) |
| 228 |
let err = vh * f64::from_bits(0x3cc0000000000000); |
| 229 |
|
| 230 |
let (p, err) = asin_eval(DoubleDouble::new(0.0, u), err); |
| 231 |
|
| 232 |
// Perform computations in double-double arithmetic: |
| 233 |
// asin(x) = pi/2 - (v_hi + v_lo) * (ASIN_COEFFS[idx][0] + p) |
| 234 |
let r0 = DoubleDouble::quick_mult(DoubleDouble::new(vl, vh), p); |
| 235 |
let mut r = DoubleDouble::from_exact_add(PI_OVER_TWO.hi, -r0.hi); |
| 236 |
|
| 237 |
let mut r_lo = PI_OVER_TWO.lo - r0.lo + r.lo; |
| 238 |
|
| 239 |
let p = DoubleDouble::mult(DoubleDouble::new(r_lo, r.hi), INV_PI_DD); |
| 240 |
r_lo = p.lo; |
| 241 |
r.hi = p.hi; |
| 242 |
|
| 243 |
let (r_upper, r_lower); |
| 244 |
|
| 245 |
#[cfg(any( |
| 246 |
all( |
| 247 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 248 |
target_feature = "fma" |
| 249 |
), |
| 250 |
target_arch = "aarch64" |
| 251 |
))] |
| 252 |
{ |
| 253 |
r_upper = f_fmla(r.hi, x_sign, f_fmla(r_lo, x_sign, err)); |
| 254 |
r_lower = f_fmla(r.hi, x_sign, f_fmla(r_lo, x_sign, -err)); |
| 255 |
} |
| 256 |
#[cfg(not(any( |
| 257 |
all( |
| 258 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 259 |
target_feature = "fma" |
| 260 |
), |
| 261 |
target_arch = "aarch64" |
| 262 |
)))] |
| 263 |
{ |
| 264 |
let r_lo = r_lo * x_sign; |
| 265 |
let r_hi = r.hi * x_sign; |
| 266 |
r_upper = r_hi + (r_lo + err); |
| 267 |
r_lower = r.hi + (r_lo - err); |
| 268 |
} |
| 269 |
|
| 270 |
if r_upper == r_lower { |
| 271 |
return r_upper; |
| 272 |
} |
| 273 |
|
| 274 |
// Ziv's accuracy test failed, we redo the computations in Float128. |
| 275 |
// Recalculate mod 1/64. |
| 276 |
let idx = (u * f64::from_bits(0x4050000000000000)).cpu_round() as usize; |
| 277 |
|
| 278 |
// After the first step of Newton-Raphson approximating v = sqrt(u), we have |
| 279 |
// that: |
| 280 |
// sqrt(u) = v_hi + h / (sqrt(u) + v_hi) |
| 281 |
// v_lo = h / (2 * v_hi) |
| 282 |
// With error: |
| 283 |
// sqrt(u) - (v_hi + v_lo) = h * ( 1/(sqrt(u) + v_hi) - 1/(2*v_hi) ) |
| 284 |
// = -h^2 / (2*v * (sqrt(u) + v)^2). |
| 285 |
// Since: |
| 286 |
// (sqrt(u) + v_hi)^2 ~ (2sqrt(u))^2 = 4u, |
| 287 |
// we can add another correction term to (v_hi + v_lo) that is: |
| 288 |
// v_ll = -h^2 / (2*v_hi * 4u) |
| 289 |
// = -v_lo * (h / 4u) |
| 290 |
// = -vl * (h / 8u), |
| 291 |
// making the errors: |
| 292 |
// sqrt(u) - (v_hi + v_lo + v_ll) = O(h^3) |
| 293 |
// well beyond 128-bit precision needed. |
| 294 |
|
| 295 |
// Get the rounding error of vl = 2 * v_lo ~ h / vh |
| 296 |
// Get full product of vh * vl |
| 297 |
let vl_lo; |
| 298 |
#[cfg(any( |
| 299 |
all( |
| 300 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 301 |
target_feature = "fma" |
| 302 |
), |
| 303 |
target_arch = "aarch64" |
| 304 |
))] |
| 305 |
{ |
| 306 |
vl_lo = f_fmla(-v_hi, vl, h) / v_hi; |
| 307 |
} |
| 308 |
#[cfg(not(any( |
| 309 |
all( |
| 310 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 311 |
target_feature = "fma" |
| 312 |
), |
| 313 |
target_arch = "aarch64" |
| 314 |
)))] |
| 315 |
{ |
| 316 |
let vh_vl = DoubleDouble::from_exact_mult(v_hi, vl); |
| 317 |
vl_lo = ((h - vh_vl.hi) - vh_vl.lo) / v_hi; |
| 318 |
} |
| 319 |
|
| 320 |
// vll = 2*v_ll = -vl * (h / (4u)). |
| 321 |
let t = h * (-0.25) / u; |
| 322 |
let vll = f_fmla(vl, t, vl_lo); |
| 323 |
// m_v = -(v_hi + v_lo + v_ll). |
| 324 |
let mv0 = DyadicFloat128::new_from_f64(vl) + DyadicFloat128::new_from_f64(vll); |
| 325 |
let mut m_v = DyadicFloat128::new_from_f64(vh) + mv0; |
| 326 |
m_v.sign = DyadicSign::Neg; |
| 327 |
|
| 328 |
// Perform computations in Float128: |
| 329 |
// asin(x) = pi/2 - (v_hi + v_lo + vll) * P(u). |
| 330 |
let y_f128 = |
| 331 |
DyadicFloat128::new_from_f64(f_fmla(idx as f64, f64::from_bits(0xbf90000000000000), u)); |
| 332 |
|
| 333 |
const PI_OVER_TWO_F128: DyadicFloat128 = DyadicFloat128 { |
| 334 |
sign: DyadicSign::Pos, |
| 335 |
exponent: -127, |
| 336 |
mantissa: 0xc90fdaa2_2168c234_c4c6628b_80dc1cd1_u128, |
| 337 |
}; |
| 338 |
|
| 339 |
let p_f128 = asin_eval_dyadic(y_f128, idx); |
| 340 |
let r0_f128 = m_v * p_f128; |
| 341 |
let mut r_f128 = PI_OVER_TWO_F128 + r0_f128; |
| 342 |
|
| 343 |
if x.is_sign_negative() { |
| 344 |
r_f128.sign = DyadicSign::Neg; |
| 345 |
} |
| 346 |
|
| 347 |
r_f128 = r_f128.quick_mul(&crate::acospi::INV_PI_F128); |
| 348 |
|
| 349 |
r_f128.fast_as_f64() |
| 350 |
} |
| 351 |
|
| 352 |
#[cfg(test)] |
| 353 |
mod tests { |
| 354 |
use super::*; |
| 355 |
|
| 356 |
#[test] |
| 357 |
fn f_asinpi_test() { |
| 358 |
assert_eq!( |
| 359 |
f_asinpi(-0.00000000032681723993732703), |
| 360 |
-0.00000000010402915844735117 |
| 361 |
); |
| 362 |
assert_eq!(f_asinpi(0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000017801371778309684), 0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000005666352624669099); |
| 363 |
assert_eq!(f_asinpi(0.000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000026752519513526076), 0.000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000008515591441480124); |
| 364 |
assert_eq!(f_asinpi(-0.4), -0.13098988043445461); |
| 365 |
assert_eq!(f_asinpi(-0.8), -0.2951672353008666); |
| 366 |
assert_eq!(f_asinpi(0.4332432142124432), 0.14263088583055605); |
| 367 |
assert_eq!(f_asinpi(0.8543543534343434), 0.326047108714517); |
| 368 |
assert_eq!(f_asinpi(0.00323146509843243), 0.0010286090778797426); |
| 369 |
} |
| 370 |
} |