| 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::bits::EXP_MASK; |
| 30 |
use crate::common::{dyad_fmla, f_fmla}; |
| 31 |
use crate::double_double::DoubleDouble; |
| 32 |
use crate::sin::range_reduction_small; |
| 33 |
use crate::sincos_reduce::LargeArgumentReduction; |
| 34 |
use crate::tangent::tanpi_table::TAN_K_PI_OVER_128; |
| 35 |
|
| 36 |
#[inline] |
| 37 |
pub(crate) fn tan_eval(u: DoubleDouble) -> (DoubleDouble, f64) { |
| 38 |
// Evaluate tan(y) = tan(x - k * (pi/128)) |
| 39 |
// We use the degree-9 Taylor approximation: |
| 40 |
// tan(y) ~ P(y) = y + y^3/3 + 2*y^5/15 + 17*y^7/315 + 62*y^9/2835 |
| 41 |
// Then the error is bounded by: |
| 42 |
// |tan(y) - P(y)| < 2^-6 * |y|^11 < 2^-6 * 2^-66 = 2^-72. |
| 43 |
// For y ~ u_hi + u_lo, fully expanding the polynomial and drop any terms |
| 44 |
// < ulp(u_hi^3) gives us: |
| 45 |
// P(y) = y + y^3/3 + 2*y^5/15 + 17*y^7/315 + 62*y^9/2835 = ... |
| 46 |
// ~ u_hi + u_hi^3 * (1/3 + u_hi^2 * (2/15 + u_hi^2 * (17/315 + |
| 47 |
// + u_hi^2 * 62/2835))) + |
| 48 |
// + u_lo (1 + u_hi^2 * (1 + u_hi^2 * 2/3)) |
| 49 |
let u_hi_sq = u.hi * u.hi; // Error < ulp(u_hi^2) < 2^(-6 - 52) = 2^-58. |
| 50 |
// p1 ~ 17/315 + u_hi^2 62 / 2835. |
| 51 |
let p1 = f_fmla( |
| 52 |
u_hi_sq, |
| 53 |
f64::from_bits(0x3f9664f4882c10fa), |
| 54 |
f64::from_bits(0x3faba1ba1ba1ba1c), |
| 55 |
); |
| 56 |
// p2 ~ 1/3 + u_hi^2 2 / 15. |
| 57 |
let p2 = f_fmla( |
| 58 |
u_hi_sq, |
| 59 |
f64::from_bits(0x3fc1111111111111), |
| 60 |
f64::from_bits(0x3fd5555555555555), |
| 61 |
); |
| 62 |
// q1 ~ 1 + u_hi^2 * 2/3. |
| 63 |
let q1 = f_fmla(u_hi_sq, f64::from_bits(0x3fe5555555555555), 1.0); |
| 64 |
let u_hi_3 = u_hi_sq * u.hi; |
| 65 |
let u_hi_4 = u_hi_sq * u_hi_sq; |
| 66 |
// p3 ~ 1/3 + u_hi^2 * (2/15 + u_hi^2 * (17/315 + u_hi^2 * 62/2835)) |
| 67 |
let p3 = f_fmla(u_hi_4, p1, p2); |
| 68 |
// q2 ~ 1 + u_hi^2 * (1 + u_hi^2 * 2/3) |
| 69 |
let q2 = f_fmla(u_hi_sq, q1, 1.0); |
| 70 |
let tan_lo = f_fmla(u_hi_3, p3, u.lo * q2); |
| 71 |
// Overall, |tan(y) - (u_hi + tan_lo)| < ulp(u_hi^3) <= 2^-71. |
| 72 |
// And the relative errors is: |
| 73 |
// |(tan(y) - (u_hi + tan_lo)) / tan(y) | <= 2*ulp(u_hi^2) < 2^-64 |
| 74 |
let err = f_fmla( |
| 75 |
u_hi_3.abs(), |
| 76 |
f64::from_bits(0x3cc0000000000000), |
| 77 |
f64::from_bits(0x3990000000000000), |
| 78 |
); |
| 79 |
(DoubleDouble::from_exact_add(u.hi, tan_lo), err) |
| 80 |
} |
| 81 |
|
| 82 |
#[inline] |
| 83 |
pub(crate) fn tan_eval_dd(x: DoubleDouble) -> DoubleDouble { |
| 84 |
// Generated by Sollya: |
| 85 |
// d = [0, pi/128]; |
| 86 |
// f_tan = tan(x)/x; |
| 87 |
// Q = fpminimax(f_tan, [|0, 2, 4, 6, 8, 10, 12|], [|107...|], d); |
| 88 |
const C: [(u64, u64); 7] = [ |
| 89 |
(0x0000000000000000, 0x3ff0000000000000), |
| 90 |
(0x3c75555549d35a4b, 0x3fd5555555555555), |
| 91 |
(0x3c413dd6ea580288, 0x3fc1111111111111), |
| 92 |
(0x3c4e100b946f0c89, 0x3faba1ba1ba1b9fe), |
| 93 |
(0xbc3c180dfac2b8c3, 0x3f9664f488307189), |
| 94 |
(0x3c1fd691c256a14a, 0x3f8226e300c1988e), |
| 95 |
(0x3bec7b08c90fdfc0, 0x3f6d739baeacd204), |
| 96 |
]; |
| 97 |
let x2 = DoubleDouble::quick_mult(x, x); |
| 98 |
let mut p = DoubleDouble::quick_mul_add( |
| 99 |
x2, |
| 100 |
DoubleDouble::from_bit_pair(C[6]), |
| 101 |
DoubleDouble::from_bit_pair(C[5]), |
| 102 |
); |
| 103 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[4])); |
| 104 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); |
| 105 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); |
| 106 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); |
| 107 |
p = DoubleDouble::quick_mul_add_f64(x2, p, f64::from_bits(0x3ff0000000000000)); |
| 108 |
let r = DoubleDouble::quick_mult(p, x); |
| 109 |
DoubleDouble::from_exact_add(r.hi, r.lo) |
| 110 |
} |
| 111 |
|
| 112 |
#[cold] |
| 113 |
fn tan_accurate(y: DoubleDouble, tan_k: DoubleDouble) -> f64 { |
| 114 |
// Computes tan(x) through identities |
| 115 |
// tan(a+b) = (tan(a) + tan(b)) / (1 - tan(a)tan(b)) = (tan(y) + tan(k*pi/128)) / (1 - tan(y)*tan(k*pi/128)) |
| 116 |
let tan_y = tan_eval_dd(y); |
| 117 |
|
| 118 |
// num = tan(y) + tan(k*pi/64) |
| 119 |
let num_dd = DoubleDouble::full_dd_add(tan_y, tan_k); |
| 120 |
// den = 1 - tan(y)*tan(k*pi/64) |
| 121 |
let den_dd = DoubleDouble::mul_add_f64(tan_y, -tan_k, 1.); |
| 122 |
|
| 123 |
let tan_x = DoubleDouble::div(num_dd, den_dd); |
| 124 |
tan_x.to_f64() |
| 125 |
} |
| 126 |
|
| 127 |
#[cold] |
| 128 |
fn tan_near_zero_hard(x: f64) -> DoubleDouble { |
| 129 |
// Generated by Sollya: |
| 130 |
// d = [0, 0.03125]; |
| 131 |
// f_tan = tan(x)/x; |
| 132 |
// Q = fpminimax(f_tan, [|0, 2, 4, 6, 8, 10, 12|], [|1, 107...|], d); |
| 133 |
const C: [(u64, u64); 7] = [ |
| 134 |
(0x0000000000000000, 0x3ff0000000000000), |
| 135 |
(0x3c7555548455da94, 0x3fd5555555555555), |
| 136 |
(0x3c4306a6358cc810, 0x3fc1111111111111), |
| 137 |
(0x3c2ca9cd025ea98c, 0x3faba1ba1ba1b952), |
| 138 |
(0x3c3cb012803c55a5, 0x3f9664f4883eb962), |
| 139 |
(0x3c28afc1adb8f202, 0x3f8226e276097461), |
| 140 |
(0xbbdf8f911392f348, 0x3f6d7791601277ea), |
| 141 |
]; |
| 142 |
let x2 = DoubleDouble::from_exact_mult(x, x); |
| 143 |
let mut p = DoubleDouble::quick_mul_add( |
| 144 |
x2, |
| 145 |
DoubleDouble::from_bit_pair(C[6]), |
| 146 |
DoubleDouble::from_bit_pair(C[5]), |
| 147 |
); |
| 148 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[4])); |
| 149 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); |
| 150 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); |
| 151 |
p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); |
| 152 |
p = DoubleDouble::quick_mul_add_f64(x2, p, f64::from_bits(0x3ff0000000000000)); |
| 153 |
DoubleDouble::quick_mult_f64(p, x) |
| 154 |
} |
| 155 |
|
| 156 |
/// Tangent in double precision |
| 157 |
/// |
| 158 |
/// ULP 0.5 |
| 159 |
pub fn f_tan(x: f64) -> f64 { |
| 160 |
let x_e = (x.to_bits() >> 52) & 0x7ff; |
| 161 |
const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; |
| 162 |
|
| 163 |
let y: DoubleDouble; |
| 164 |
let k; |
| 165 |
|
| 166 |
let mut argument_reduction = LargeArgumentReduction::default(); |
| 167 |
|
| 168 |
if x_e < E_BIAS + 16 { |
| 169 |
// |x| < 2^16 |
| 170 |
if x_e < E_BIAS - 5 { |
| 171 |
// |x| < 2^-5 |
| 172 |
if x_e < E_BIAS - 27 { |
| 173 |
// |x| < 2^-27, |tan(x) - x| < ulp(x)/2. |
| 174 |
if x == 0.0 { |
| 175 |
// Signed zeros. |
| 176 |
return x + x; |
| 177 |
} |
| 178 |
return dyad_fmla(x, f64::from_bits(0x3c90000000000000), x); |
| 179 |
} |
| 180 |
let x2 = x * x; |
| 181 |
let x4 = x2 * x2; |
| 182 |
// Generated by Sollya: |
| 183 |
// d = [0, 0.03125]; |
| 184 |
// f_tan = tan(x)/x; |
| 185 |
// Q = fpminimax(f_tan, [|0, 2, 4, 6, 8|], [|D...|], d); |
| 186 |
const C: [u64; 4] = [ |
| 187 |
0x3fd555555555554b, |
| 188 |
0x3fc1111111142d5b, |
| 189 |
0x3faba1b9860ed843, |
| 190 |
0x3f966a802210f5bb, |
| 191 |
]; |
| 192 |
let p01 = f_fmla(x2, f64::from_bits(C[1]), f64::from_bits(C[0])); |
| 193 |
let p23 = f_fmla(x2, f64::from_bits(C[3]), f64::from_bits(C[2])); |
| 194 |
let w0 = f_fmla(x4, p23, p01); |
| 195 |
let w1 = x2 * w0 * x; |
| 196 |
let r = DoubleDouble::from_exact_add(x, w1); |
| 197 |
let err = f_fmla( |
| 198 |
x2, |
| 199 |
f64::from_bits(0x3cb0000000000000), // 2^-52 |
| 200 |
f64::from_bits(0x3be0000000000000), // 2^-65 |
| 201 |
); |
| 202 |
let ub = r.hi + (r.lo + err); |
| 203 |
let lb = r.hi + (r.lo - err); |
| 204 |
if ub == lb { |
| 205 |
return ub; |
| 206 |
} |
| 207 |
return tan_near_zero_hard(x).to_f64(); |
| 208 |
} else { |
| 209 |
// Small range reduction. |
| 210 |
(y, k) = range_reduction_small(x); |
| 211 |
} |
| 212 |
} else { |
| 213 |
// Inf or NaN |
| 214 |
if x_e > 2 * E_BIAS { |
| 215 |
if x.is_nan() { |
| 216 |
return f64::NAN; |
| 217 |
} |
| 218 |
// tan(+-Inf) = NaN |
| 219 |
return x + f64::NAN; |
| 220 |
} |
| 221 |
|
| 222 |
// Large range reduction. |
| 223 |
(k, y) = argument_reduction.reduce(x); |
| 224 |
} |
| 225 |
|
| 226 |
let (tan_y, err) = tan_eval(y); |
| 227 |
|
| 228 |
// Computes tan(x) through identities. |
| 229 |
// tan(a+b) = (tan(a) + tan(b)) / (1 - tan(a)tan(b)) = (tan(y) + tan(k*pi/128)) / (1 - tan(y)*tan(k*pi/128)) |
| 230 |
let tan_k = DoubleDouble::from_bit_pair(TAN_K_PI_OVER_128[(k & 255) as usize]); |
| 231 |
|
| 232 |
// num = tan(y) + tan(k*pi/64) |
| 233 |
let num_dd = DoubleDouble::add(tan_y, tan_k); |
| 234 |
// den = 1 - tan(y)*tan(k*pi/64) |
| 235 |
let den_dd = DoubleDouble::mul_add_f64(tan_y, -tan_k, 1.); |
| 236 |
|
| 237 |
let tan_x = DoubleDouble::div(num_dd, den_dd); |
| 238 |
|
| 239 |
// Simple error bound: |1 / den_dd| < 2^(1 + floor(-log2(den_dd)))). |
| 240 |
let den_inv = ((E_BIAS + 1) << (52 + 1)) - (den_dd.hi.to_bits() & EXP_MASK); |
| 241 |
// For tan_x = (num_dd + err) / (den_dd + err), the error is bounded by: |
| 242 |
// | tan_x - num_dd / den_dd | <= err * ( 1 + | tan_x * den_dd | ). |
| 243 |
let tan_err = err * f_fmla(f64::from_bits(den_inv), tan_x.hi.abs(), 1.0); |
| 244 |
|
| 245 |
let err_higher = tan_x.lo + tan_err; |
| 246 |
let err_lower = tan_x.lo - tan_err; |
| 247 |
|
| 248 |
let tan_upper = tan_x.hi + err_higher; |
| 249 |
let tan_lower = tan_x.hi + err_lower; |
| 250 |
|
| 251 |
// Ziv_s rounding test. |
| 252 |
if tan_upper == tan_lower { |
| 253 |
return tan_upper; |
| 254 |
} |
| 255 |
|
| 256 |
tan_accurate(y, tan_k) |
| 257 |
} |
| 258 |
|
| 259 |
#[cfg(test)] |
| 260 |
mod tests { |
| 261 |
use super::*; |
| 262 |
|
| 263 |
#[test] |
| 264 |
fn tan_test() { |
| 265 |
assert_eq!(f_tan(0.0003242153424), 0.0003242153537600293); |
| 266 |
assert_eq!(f_tan(-0.3242153424), -0.3360742441422686); |
| 267 |
assert_eq!(f_tan(0.3242153424), 0.3360742441422686); |
| 268 |
assert_eq!(f_tan(-97301943219974110.), 0.000000000000000481917514979303); |
| 269 |
assert_eq!(f_tan(0.0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000007291122019556397), |
| 270 |
0.0000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000007291122019556397); |
| 271 |
assert_eq!(f_tan(0.0), 0.0); |
| 272 |
assert_eq!(f_tan(0.4324324324), 0.4615683710506729); |
| 273 |
assert_eq!(f_tan(1.0), 1.5574077246549023); |
| 274 |
assert_eq!(f_tan(-0.5), -0.5463024898437905); |
| 275 |
assert!(f_tan(f64::INFINITY).is_nan()); |
| 276 |
assert!(f_tan(f64::NEG_INFINITY).is_nan()); |
| 277 |
assert!(f_tan(f64::NAN).is_nan()); |
| 278 |
} |
| 279 |
} |