| 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::common::{dyad_fmla, f_fmla, min_normal_f64}; |
| 30 |
use crate::double_double::DoubleDouble; |
| 31 |
use crate::dyadic_float::{DyadicFloat128, DyadicSign}; |
| 32 |
use crate::rounding::CpuRound; |
| 33 |
use crate::sin_helper::sincos_eval_dd; |
| 34 |
use crate::sin_table::SIN_K_PI_OVER_128; |
| 35 |
use crate::sincos_dyadic::SIN_K_PI_OVER_128_F128; |
| 36 |
use crate::sincos_reduce::LargeArgumentReduction; |
| 37 |
|
| 38 |
// For 2^-7 < |x| < 2^16, return k and u such that: |
| 39 |
// k = round(x * 128/pi) |
| 40 |
// x mod pi/128 = x - k * pi/128 ~ u.hi + u.lo |
| 41 |
// Error bound: |
| 42 |
// |(x - k * pi/128) - (u_hi + u_lo)| <= max(ulp(ulp(u_hi)), 2^-119) |
| 43 |
// <= 2^-111. |
| 44 |
#[inline] |
| 45 |
pub(crate) fn range_reduction_small(x: f64) -> (DoubleDouble, u64) { |
| 46 |
const MPI_OVER_128: [u64; 3] = [0xbf9921fb54400000, 0xbd70b4611a600000, 0xbb43198a2e037073]; |
| 47 |
const ONE_TWENTY_EIGHT_OVER_PI_D: f64 = f64::from_bits(0x40445f306dc9c883); |
| 48 |
let prod_hi = x * ONE_TWENTY_EIGHT_OVER_PI_D; |
| 49 |
let kd = prod_hi.cpu_round(); |
| 50 |
|
| 51 |
// Let y = x - k * (pi/128) |
| 52 |
// Then |y| < pi / 256 |
| 53 |
// With extra rounding errors, we can bound |y| < 1.6 * 2^-7. |
| 54 |
let y_hi = f_fmla(kd, f64::from_bits(MPI_OVER_128[0]), x); // Exact |
| 55 |
// |u.hi| < 1.6*2^-7 |
| 56 |
let u_hi = f_fmla(kd, f64::from_bits(MPI_OVER_128[1]), y_hi); |
| 57 |
|
| 58 |
let u0 = y_hi - u_hi; // Exact |
| 59 |
// |u.lo| <= max(ulp(u.hi), |kd * MPI_OVER_128[2]|) |
| 60 |
let u1 = f_fmla(kd, f64::from_bits(MPI_OVER_128[1]), u0); // Exact |
| 61 |
let u_lo = f_fmla(kd, f64::from_bits(MPI_OVER_128[2]), u1); |
| 62 |
// Error bound: |
| 63 |
// |x - k * pi/128| - (u.hi + u.lo) <= ulp(u.lo) |
| 64 |
// <= ulp(max(ulp(u.hi), kd*MPI_OVER_128[2])) |
| 65 |
// <= 2^(-7 - 104) = 2^-111. |
| 66 |
(DoubleDouble::new(u_lo, u_hi), unsafe { |
| 67 |
kd.to_int_unchecked::<i64>() as u64 // indeterminate values is always filtered out before this call, as well only lowest bits are used |
| 68 |
}) |
| 69 |
} |
| 70 |
|
| 71 |
#[inline] |
| 72 |
pub(crate) fn get_sin_k_rational(kk: u64) -> DyadicFloat128 { |
| 73 |
let idx = if (kk & 64) != 0 { |
| 74 |
64 - (kk & 63) |
| 75 |
} else { |
| 76 |
kk & 63 |
| 77 |
}; |
| 78 |
let mut ans = SIN_K_PI_OVER_128_F128[idx as usize]; |
| 79 |
if (kk & 128) != 0 { |
| 80 |
ans.sign = DyadicSign::Neg; |
| 81 |
} |
| 82 |
ans |
| 83 |
} |
| 84 |
|
| 85 |
pub(crate) struct SinCos { |
| 86 |
pub(crate) v_sin: DoubleDouble, |
| 87 |
pub(crate) v_cos: DoubleDouble, |
| 88 |
pub(crate) err: f64, |
| 89 |
} |
| 90 |
|
| 91 |
#[inline] |
| 92 |
pub(crate) fn sincos_eval(u: DoubleDouble) -> SinCos { |
| 93 |
// Evaluate sin(y) = sin(x - k * (pi/128)) |
| 94 |
// We use the degree-7 Taylor approximation: |
| 95 |
// sin(y) ~ y - y^3/3! + y^5/5! - y^7/7! |
| 96 |
// Then the error is bounded by: |
| 97 |
// |sin(y) - (y - y^3/3! + y^5/5! - y^7/7!)| < |y|^9/9! < 2^-54/9! < 2^-72. |
| 98 |
// For y ~ u_hi + u_lo, fully expanding the polynomial and drop any terms |
| 99 |
// < ulp(u_hi^3) gives us: |
| 100 |
// y - y^3/3! + y^5/5! - y^7/7! = ... |
| 101 |
// ~ u_hi + u_hi^3 * (-1/6 + u_hi^2 * (1/120 - u_hi^2 * 1/5040)) + |
| 102 |
// + u_lo (1 + u_hi^2 * (-1/2 + u_hi^2 / 24)) |
| 103 |
let u_hi_sq = u.hi * u.hi; // Error < ulp(u_hi^2) < 2^(-6 - 52) = 2^-58. |
| 104 |
// p1 ~ 1/120 + u_hi^2 / 5040. |
| 105 |
let p1 = f_fmla( |
| 106 |
u_hi_sq, |
| 107 |
f64::from_bits(0xbf2a01a01a01a01a), |
| 108 |
f64::from_bits(0x3f81111111111111), |
| 109 |
); |
| 110 |
// q1 ~ -1/2 + u_hi^2 / 24. |
| 111 |
let q1 = f_fmla( |
| 112 |
u_hi_sq, |
| 113 |
f64::from_bits(0x3fa5555555555555), |
| 114 |
f64::from_bits(0xbfe0000000000000), |
| 115 |
); |
| 116 |
let u_hi_3 = u_hi_sq * u.hi; |
| 117 |
// p2 ~ -1/6 + u_hi^2 (1/120 - u_hi^2 * 1/5040) |
| 118 |
let p2 = f_fmla(u_hi_sq, p1, f64::from_bits(0xbfc5555555555555)); |
| 119 |
// q2 ~ 1 + u_hi^2 (-1/2 + u_hi^2 / 24) |
| 120 |
let q2 = f_fmla(u_hi_sq, q1, 1.0); |
| 121 |
let sin_lo = f_fmla(u_hi_3, p2, u.lo * q2); |
| 122 |
// Overall, |sin(y) - (u_hi + sin_lo)| < 2*ulp(u_hi^3) < 2^-69. |
| 123 |
|
| 124 |
// Evaluate cos(y) = cos(x - k * (pi/128)) |
| 125 |
// We use the degree-8 Taylor approximation: |
| 126 |
// cos(y) ~ 1 - y^2/2 + y^4/4! - y^6/6! + y^8/8! |
| 127 |
// Then the error is bounded by: |
| 128 |
// |cos(y) - (...)| < |y|^10/10! < 2^-81 |
| 129 |
// For y ~ u_hi + u_lo, fully expanding the polynomial and drop any terms |
| 130 |
// < ulp(u_hi^3) gives us: |
| 131 |
// 1 - y^2/2 + y^4/4! - y^6/6! + y^8/8! = ... |
| 132 |
// ~ 1 - u_hi^2/2 + u_hi^4(1/24 + u_hi^2 (-1/720 + u_hi^2/40320)) + |
| 133 |
// + u_hi u_lo (-1 + u_hi^2/6) |
| 134 |
// We compute 1 - u_hi^2 accurately: |
| 135 |
// v_hi + v_lo ~ 1 - u_hi^2/2 |
| 136 |
// with error <= 2^-105. |
| 137 |
let u_hi_neg_half = (-0.5) * u.hi; |
| 138 |
|
| 139 |
let (mut v_lo, v_hi); |
| 140 |
|
| 141 |
#[cfg(any( |
| 142 |
all( |
| 143 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 144 |
target_feature = "fma" |
| 145 |
), |
| 146 |
target_arch = "aarch64" |
| 147 |
))] |
| 148 |
{ |
| 149 |
v_hi = f_fmla(u.hi, u_hi_neg_half, 1.0); |
| 150 |
v_lo = 1.0 - v_hi; // Exact |
| 151 |
v_lo = f_fmla(u.hi, u_hi_neg_half, v_lo); |
| 152 |
} |
| 153 |
|
| 154 |
#[cfg(not(any( |
| 155 |
all( |
| 156 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 157 |
target_feature = "fma" |
| 158 |
), |
| 159 |
target_arch = "aarch64" |
| 160 |
)))] |
| 161 |
{ |
| 162 |
let u_hi_sq_neg_half = DoubleDouble::from_exact_mult(u.hi, u_hi_neg_half); |
| 163 |
let v = DoubleDouble::from_exact_add(1.0, u_hi_sq_neg_half.hi); |
| 164 |
v_lo = v.lo; |
| 165 |
v_lo += u_hi_sq_neg_half.lo; |
| 166 |
v_hi = v.hi; |
| 167 |
} |
| 168 |
|
| 169 |
// r1 ~ -1/720 + u_hi^2 / 40320 |
| 170 |
let r1 = f_fmla( |
| 171 |
u_hi_sq, |
| 172 |
f64::from_bits(0x3efa01a01a01a01a), |
| 173 |
f64::from_bits(0xbf56c16c16c16c17), |
| 174 |
); |
| 175 |
// s1 ~ -1 + u_hi^2 / 6 |
| 176 |
let s1 = f_fmla(u_hi_sq, f64::from_bits(0x3fc5555555555555), -1.0); |
| 177 |
let u_hi_4 = u_hi_sq * u_hi_sq; |
| 178 |
let u_hi_u_lo = u.hi * u.lo; |
| 179 |
// r2 ~ 1/24 + u_hi^2 (-1/720 + u_hi^2 / 40320) |
| 180 |
let r2 = f_fmla(u_hi_sq, r1, f64::from_bits(0x3fa5555555555555)); |
| 181 |
// s2 ~ v_lo + u_hi * u_lo * (-1 + u_hi^2 / 6) |
| 182 |
let s2 = f_fmla(u_hi_u_lo, s1, v_lo); |
| 183 |
let cos_lo = f_fmla(u_hi_4, r2, s2); |
| 184 |
// Overall, |cos(y) - (v_hi + cos_lo)| < 2*ulp(u_hi^4) < 2^-75. |
| 185 |
|
| 186 |
let sin_u = DoubleDouble::from_exact_add(u.hi, sin_lo); |
| 187 |
let cos_u = DoubleDouble::from_exact_add(v_hi, cos_lo); |
| 188 |
|
| 189 |
let err = f_fmla( |
| 190 |
u_hi_3, |
| 191 |
f64::from_bits(0x3cc0000000000000), |
| 192 |
f64::from_bits(0x3960000000000000), |
| 193 |
); |
| 194 |
|
| 195 |
SinCos { |
| 196 |
v_sin: sin_u, |
| 197 |
v_cos: cos_u, |
| 198 |
err, |
| 199 |
} |
| 200 |
} |
| 201 |
|
| 202 |
#[cold] |
| 203 |
#[inline(never)] |
| 204 |
fn sin_accurate(y: DoubleDouble, sin_k: DoubleDouble, cos_k: DoubleDouble) -> f64 { |
| 205 |
let r_sincos = sincos_eval_dd(y); |
| 206 |
|
| 207 |
// k is an integer and -pi / 256 <= y <= pi / 256. |
| 208 |
// Then sin(x) = sin((k * pi/128 + y) |
| 209 |
// = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128) |
| 210 |
|
| 211 |
let sin_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, sin_k); |
| 212 |
let cos_k_sin_y = DoubleDouble::quick_mult(r_sincos.v_sin, cos_k); |
| 213 |
|
| 214 |
let rr = DoubleDouble::full_dd_add(sin_k_cos_y, cos_k_sin_y); |
| 215 |
rr.to_f64() |
| 216 |
} |
| 217 |
|
| 218 |
#[cold] |
| 219 |
pub(crate) fn sin_accurate_near_zero(x: f64) -> f64 { |
| 220 |
// Generated by Sollya: |
| 221 |
// d = [0, 0.0490874]; |
| 222 |
// f_sin = sin(y)/y; |
| 223 |
// Q = fpminimax(f_sin, [|0, 2, 4, 6, 8, 10, 12|], [|107...|], d, relative, floating); |
| 224 |
const C: [(u64, u64); 7] = [ |
| 225 |
(0x0000000000000000, 0x3ff0000000000000), |
| 226 |
(0xbc65555555555217, 0xbfc5555555555555), |
| 227 |
(0x3c011110f7d49e8c, 0x3f81111111111111), |
| 228 |
(0xbb65e5b30986fc80, 0xbf2a01a01a01a01a), |
| 229 |
(0xbb689e86245bd566, 0x3ec71de3a556c6e5), |
| 230 |
(0xbaccb3d55ccfca58, 0xbe5ae64567954935), |
| 231 |
(0x3a6333ef7a00ce40, 0x3de6120ca00ce3a2), |
| 232 |
]; |
| 233 |
let x2 = DoubleDouble::from_exact_mult(x, x); |
| 234 |
let mut p = DoubleDouble::quick_mul_add( |
| 235 |
x2, |
| 236 |
DoubleDouble::from_bit_pair(C[6]), |
| 237 |
DoubleDouble::from_bit_pair(C[5]), |
| 238 |
); |
| 239 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[4])); |
| 240 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); |
| 241 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); |
| 242 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); |
| 243 |
p = DoubleDouble::quick_mul_add_f64(x2, p, f64::from_bits(0x3ff0000000000000)); |
| 244 |
p = DoubleDouble::quick_mult_f64(p, x); |
| 245 |
p.to_f64() |
| 246 |
} |
| 247 |
|
| 248 |
/// Sine for double precision |
| 249 |
/// |
| 250 |
/// ULP 0.5 |
| 251 |
pub fn f_sin(x: f64) -> f64 { |
| 252 |
let x_e = (x.to_bits() >> 52) & 0x7ff; |
| 253 |
const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; |
| 254 |
|
| 255 |
let y: DoubleDouble; |
| 256 |
let k; |
| 257 |
|
| 258 |
let mut argument_reduction = LargeArgumentReduction::default(); |
| 259 |
|
| 260 |
if x_e < E_BIAS + 16 { |
| 261 |
// |x| < 2^32 (with FMA) or |x| < 2^23 (w/o FMA) |
| 262 |
let ax = x.to_bits() & 0x7fff_ffff_ffff_ffff; |
| 263 |
if ax <= 0x3fa921fbd34a9550 { |
| 264 |
// |x| <= 0.0490874 |
| 265 |
if x_e < E_BIAS - 26 { |
| 266 |
// |x| < 2^-26 |
| 267 |
if x == 0.0 { |
| 268 |
// Signed zeros. |
| 269 |
return x; |
| 270 |
} |
| 271 |
|
| 272 |
// For |x| < 2^-26, |sin(x) - x| < ulp(x)/2. |
| 273 |
return dyad_fmla(x, f64::from_bits(0xbc90000000000000), x); |
| 274 |
} |
| 275 |
|
| 276 |
// Polynomial for sin(x)/x |
| 277 |
// Generated by Sollya: |
| 278 |
// d = [0, 0.0490874]; |
| 279 |
// f_sin = sin(y)/y; |
| 280 |
// Q = fpminimax(f_sin, [|0, 2, 4, 6, 8|], [|1, D...|], d, relative, floating); |
| 281 |
|
| 282 |
let x2 = x * x; |
| 283 |
let x4 = x2 * x2; |
| 284 |
const C: [u64; 4] = [ |
| 285 |
0xbfc5555555555555, |
| 286 |
0x3f8111111110e45a, |
| 287 |
0xbf2a019ffd7fdaaf, |
| 288 |
0x3ec71819b9bf01ef, |
| 289 |
]; |
| 290 |
let p01 = f_fmla(x2, f64::from_bits(C[1]), f64::from_bits(C[0])); |
| 291 |
let p23 = f_fmla(x2, f64::from_bits(C[3]), f64::from_bits(C[2])); |
| 292 |
let w0 = f_fmla(x4, p23, p01); |
| 293 |
let w1 = x2 * w0 * x; |
| 294 |
let r = DoubleDouble::from_exact_add(x, w1); |
| 295 |
let err = f_fmla( |
| 296 |
x2, |
| 297 |
f64::from_bits(0x3cb0000000000000), // 2^-52 |
| 298 |
f64::from_bits(0x3be0000000000000), // 2^-65 |
| 299 |
); |
| 300 |
let ub = r.hi + (r.lo + err); |
| 301 |
let lb = r.hi + (r.lo - err); |
| 302 |
if ub == lb { |
| 303 |
return ub; |
| 304 |
} |
| 305 |
return sin_accurate_near_zero(x); |
| 306 |
} |
| 307 |
|
| 308 |
// Small range reduction. |
| 309 |
(y, k) = range_reduction_small(x); |
| 310 |
} else { |
| 311 |
// Inf or NaN |
| 312 |
if x_e > 2 * E_BIAS { |
| 313 |
// sin(+-Inf) = NaN |
| 314 |
return x + f64::NAN; |
| 315 |
} |
| 316 |
|
| 317 |
// Large range reduction. |
| 318 |
(k, y) = argument_reduction.reduce(x); |
| 319 |
} |
| 320 |
|
| 321 |
let r_sincos = sincos_eval(y); |
| 322 |
|
| 323 |
// cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128). |
| 324 |
let sk = SIN_K_PI_OVER_128[(k & 255) as usize]; |
| 325 |
let ck = SIN_K_PI_OVER_128[((k.wrapping_add(64)) & 255) as usize]; |
| 326 |
|
| 327 |
let sin_k = DoubleDouble::from_bit_pair(sk); |
| 328 |
let cos_k = DoubleDouble::from_bit_pair(ck); |
| 329 |
|
| 330 |
let sin_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, sin_k); |
| 331 |
let cos_k_sin_y = DoubleDouble::quick_mult(r_sincos.v_sin, cos_k); |
| 332 |
|
| 333 |
// sin_k_cos_y is always >> cos_k_sin_y |
| 334 |
let mut rr = DoubleDouble::from_exact_add(sin_k_cos_y.hi, cos_k_sin_y.hi); |
| 335 |
rr.lo += sin_k_cos_y.lo + cos_k_sin_y.lo; |
| 336 |
|
| 337 |
let rlp = rr.lo + r_sincos.err; |
| 338 |
let rlm = rr.lo - r_sincos.err; |
| 339 |
|
| 340 |
let r_upper = rr.hi + rlp; // (rr.lo + ERR); |
| 341 |
let r_lower = rr.hi + rlm; // (rr.lo - ERR); |
| 342 |
|
| 343 |
// Ziv's accuracy test |
| 344 |
if r_upper == r_lower { |
| 345 |
return rr.to_f64(); |
| 346 |
} |
| 347 |
|
| 348 |
sin_accurate(y, sin_k, cos_k) |
| 349 |
} |
| 350 |
|
| 351 |
#[cold] |
| 352 |
#[inline(never)] |
| 353 |
fn cos_accurate(y: DoubleDouble, msin_k: DoubleDouble, cos_k: DoubleDouble) -> f64 { |
| 354 |
let r_sincos = sincos_eval_dd(y); |
| 355 |
|
| 356 |
// After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128). |
| 357 |
// So k is an integer and -pi / 256 <= y <= pi / 256. |
| 358 |
// Then sin(x) = sin((k * pi/128 + y) |
| 359 |
// = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128) |
| 360 |
|
| 361 |
let sin_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, cos_k); |
| 362 |
let cos_k_sin_y = DoubleDouble::quick_mult(r_sincos.v_sin, msin_k); |
| 363 |
|
| 364 |
let rr = DoubleDouble::full_dd_add(sin_k_cos_y, cos_k_sin_y); |
| 365 |
rr.to_f64() |
| 366 |
} |
| 367 |
|
| 368 |
#[cold] |
| 369 |
pub(crate) fn cos_accurate_near_zero(x: f64) -> f64 { |
| 370 |
// Generated by Sollya: |
| 371 |
// d = [0, 0.0490874]; |
| 372 |
// f_sin = sin(y)/y; |
| 373 |
// Q = fpminimax(f_sin, [|0, 2, 4, 6, 8, 10|], [|107...|], d, relative, floating); |
| 374 |
const C: [(u64, u64); 6] = [ |
| 375 |
(0xba2fa1c000000000, 0x3ff0000000000000), |
| 376 |
(0x3b1cd6ead3800000, 0xbfe0000000000000), |
| 377 |
(0x3c45112931063916, 0x3fa5555555555555), |
| 378 |
(0xbba1c1e3ab3b09e0, 0xbf56c16c16c16b2b), |
| 379 |
(0x3b937bc2f4ea7db6, 0x3efa01a019495fca), |
| 380 |
(0x3b307fd5e1b021b3, 0xbe927e0d57d894f8), |
| 381 |
]; |
| 382 |
let x2 = DoubleDouble::from_exact_mult(x, x); |
| 383 |
let mut p = DoubleDouble::quick_mul_add( |
| 384 |
x2, |
| 385 |
DoubleDouble::from_bit_pair(C[5]), |
| 386 |
DoubleDouble::from_bit_pair(C[4]), |
| 387 |
); |
| 388 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); |
| 389 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); |
| 390 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); |
| 391 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[0])); |
| 392 |
p.to_f64() |
| 393 |
} |
| 394 |
|
| 395 |
/// Cosine for double precision |
| 396 |
/// |
| 397 |
/// ULP 0.5 |
| 398 |
pub fn f_cos(x: f64) -> f64 { |
| 399 |
let x_e = (x.to_bits() >> 52) & 0x7ff; |
| 400 |
const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; |
| 401 |
|
| 402 |
let y: DoubleDouble; |
| 403 |
let k; |
| 404 |
|
| 405 |
let mut argument_reduction = LargeArgumentReduction::default(); |
| 406 |
|
| 407 |
if x_e < E_BIAS + 16 { |
| 408 |
// |x| < 2^32 (with FMA) or |x| < 2^23 (w/o FMA) |
| 409 |
let ax = x.to_bits() & 0x7fff_ffff_ffff_ffff; |
| 410 |
if ax <= 0x3fa921fbd34a9550 { |
| 411 |
// |x| <= 0.0490874 |
| 412 |
if x_e < E_BIAS - 27 { |
| 413 |
// |x| < 2^-27 |
| 414 |
if x == 0.0 { |
| 415 |
// Signed zeros. |
| 416 |
return 1.0; |
| 417 |
} |
| 418 |
// For |x| < 2^-26, |sin(x) - x| < ulp(x)/2. |
| 419 |
return 1.0 - min_normal_f64(); |
| 420 |
} |
| 421 |
|
| 422 |
// Polynomial for cos(x) |
| 423 |
// Generated by Sollya: |
| 424 |
// d = [0, 0.0490874]; |
| 425 |
// f_cos = cos(y); |
| 426 |
// Q = fpminimax(f_cos, [|0, 2, 4, 6, 8|], [|1, D...|], d, relative, floating); |
| 427 |
|
| 428 |
let x2 = x * x; |
| 429 |
let x4 = x2 * x2; |
| 430 |
const C: [u64; 4] = [ |
| 431 |
0xbfe0000000000000, |
| 432 |
0x3fa55555555554a4, |
| 433 |
0xbf56c16c1619b84a, |
| 434 |
0x3efa013d3d01cf7f, |
| 435 |
]; |
| 436 |
let p01 = f_fmla(x2, f64::from_bits(C[1]), f64::from_bits(C[0])); |
| 437 |
let p23 = f_fmla(x2, f64::from_bits(C[3]), f64::from_bits(C[2])); |
| 438 |
let w0 = f_fmla(x4, p23, p01); |
| 439 |
let w1 = x2 * w0; |
| 440 |
let r = DoubleDouble::from_exact_add(1., w1); |
| 441 |
let err = f_fmla( |
| 442 |
x2, |
| 443 |
f64::from_bits(0x3cb0000000000000), // 2^-52 |
| 444 |
f64::from_bits(0x3be0000000000000), // 2^-65 |
| 445 |
); |
| 446 |
|
| 447 |
let ub = r.hi + (r.lo + err); |
| 448 |
let lb = r.hi + (r.lo - err); |
| 449 |
if ub == lb { |
| 450 |
return ub; |
| 451 |
} |
| 452 |
return cos_accurate_near_zero(x); |
| 453 |
} else { |
| 454 |
// // Small range reduction. |
| 455 |
(y, k) = range_reduction_small(x); |
| 456 |
} |
| 457 |
} else { |
| 458 |
// Inf or NaN |
| 459 |
if x_e > 2 * E_BIAS { |
| 460 |
// cos(+-Inf) = NaN |
| 461 |
return x + f64::NAN; |
| 462 |
} |
| 463 |
|
| 464 |
// Large range reduction. |
| 465 |
(k, y) = argument_reduction.reduce(x); |
| 466 |
} |
| 467 |
let r_sincos = sincos_eval(y); |
| 468 |
|
| 469 |
// After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128). |
| 470 |
// So k is an integer and -pi / 256 <= y <= pi / 256. |
| 471 |
// Then cos(x) = cos((k * pi/128 + y) |
| 472 |
// = cos(y) * cos(k*pi/128) - sin(y) * sin(k*pi/128) |
| 473 |
let sk = SIN_K_PI_OVER_128[(k.wrapping_add(128) & 255) as usize]; |
| 474 |
let ck = SIN_K_PI_OVER_128[((k.wrapping_add(64)) & 255) as usize]; |
| 475 |
let msin_k = DoubleDouble::from_bit_pair(sk); |
| 476 |
let cos_k = DoubleDouble::from_bit_pair(ck); |
| 477 |
|
| 478 |
let cos_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, cos_k); |
| 479 |
let cos_k_msin_y = DoubleDouble::quick_mult(r_sincos.v_sin, msin_k); |
| 480 |
|
| 481 |
// cos_k_cos_y is always >> cos_k_msin_y |
| 482 |
let mut rr = DoubleDouble::from_exact_add(cos_k_cos_y.hi, cos_k_msin_y.hi); |
| 483 |
rr.lo += cos_k_cos_y.lo + cos_k_msin_y.lo; |
| 484 |
let rlp = rr.lo + r_sincos.err; |
| 485 |
let rlm = rr.lo - r_sincos.err; |
| 486 |
|
| 487 |
let r_upper = rr.hi + rlp; // (rr.lo + ERR); |
| 488 |
let r_lower = rr.hi + rlm; // (rr.lo - ERR); |
| 489 |
|
| 490 |
// Ziv's accuracy test |
| 491 |
if r_upper == r_lower { |
| 492 |
return rr.to_f64(); |
| 493 |
} |
| 494 |
cos_accurate(y, msin_k, cos_k) |
| 495 |
} |
| 496 |
|
| 497 |
#[cfg(test)] |
| 498 |
mod tests { |
| 499 |
use super::*; |
| 500 |
|
| 501 |
#[test] |
| 502 |
fn cos_test() { |
| 503 |
assert_eq!(f_cos(0.0), 1.0); |
| 504 |
assert_eq!(f_cos(0.0024312312), 0.9999970445588818); |
| 505 |
assert_eq!(f_cos(1.0), 0.5403023058681398); |
| 506 |
assert_eq!(f_cos(-0.5), 0.8775825618903728); |
| 507 |
assert!(f_cos(f64::INFINITY).is_nan()); |
| 508 |
assert!(f_cos(f64::NEG_INFINITY).is_nan()); |
| 509 |
assert!(f_cos(f64::NAN).is_nan()); |
| 510 |
} |
| 511 |
|
| 512 |
#[test] |
| 513 |
fn sin_test() { |
| 514 |
assert_eq!(f_sin(0.00000002149119332890786), 0.00000002149119332890786); |
| 515 |
assert_eq!(f_sin(2.8477476437362989E-306), 2.8477476437362989E-306); |
| 516 |
assert_eq!(f_sin(0.0024312312), 0.002431228804879309); |
| 517 |
assert_eq!(f_sin(0.0), 0.0); |
| 518 |
assert_eq!(f_sin(1.0), 0.8414709848078965); |
| 519 |
assert_eq!(f_sin(-0.5), -0.479425538604203); |
| 520 |
assert!(f_sin(f64::INFINITY).is_nan()); |
| 521 |
assert!(f_sin(f64::NEG_INFINITY).is_nan()); |
| 522 |
assert!(f_sin(f64::NAN).is_nan()); |
| 523 |
} |
| 524 |
} |