| 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::f_fmla; |
| 30 |
use crate::cube_roots::cbrtf::halley_refine_d; |
| 31 |
use crate::double_double::DoubleDouble; |
| 32 |
use crate::exponents::fast_ldexp; |
| 33 |
|
| 34 |
pub(crate) trait CbrtBackend { |
| 35 |
fn fma(&self, x: f64, y: f64, z: f64) -> f64; |
| 36 |
fn polyeval4(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64) -> f64; |
| 37 |
fn halley(&self, x: f64, a: f64) -> f64; |
| 38 |
fn exact_mul(&self, x: f64, y: f64) -> DoubleDouble; |
| 39 |
fn quick_mult_f64(&self, x: DoubleDouble, y: f64) -> DoubleDouble; |
| 40 |
} |
| 41 |
|
| 42 |
pub(crate) struct GenericCbrtBackend {} |
| 43 |
|
| 44 |
impl CbrtBackend for GenericCbrtBackend { |
| 45 |
#[inline(always)] |
| 46 |
fn fma(&self, x: f64, y: f64, z: f64) -> f64 { |
| 47 |
f_fmla(x, y, z) |
| 48 |
} |
| 49 |
#[inline(always)] |
| 50 |
fn polyeval4(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64) -> f64 { |
| 51 |
use crate::polyeval::f_polyeval4; |
| 52 |
f_polyeval4(x, a0, a1, a2, a3) |
| 53 |
} |
| 54 |
#[inline(always)] |
| 55 |
fn halley(&self, x: f64, a: f64) -> f64 { |
| 56 |
halley_refine_d(x, a) |
| 57 |
} |
| 58 |
#[inline(always)] |
| 59 |
fn exact_mul(&self, x: f64, y: f64) -> DoubleDouble { |
| 60 |
DoubleDouble::from_exact_mult(x, y) |
| 61 |
} |
| 62 |
#[inline(always)] |
| 63 |
fn quick_mult_f64(&self, x: DoubleDouble, y: f64) -> DoubleDouble { |
| 64 |
DoubleDouble::quick_mult_f64(x, y) |
| 65 |
} |
| 66 |
} |
| 67 |
|
| 68 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 69 |
pub(crate) struct FmaCbrtBackend {} |
| 70 |
|
| 71 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 72 |
impl CbrtBackend for FmaCbrtBackend { |
| 73 |
#[inline(always)] |
| 74 |
fn fma(&self, x: f64, y: f64, z: f64) -> f64 { |
| 75 |
f64::mul_add(x, y, z) |
| 76 |
} |
| 77 |
#[inline(always)] |
| 78 |
fn polyeval4(&self, x: f64, a0: f64, a1: f64, a2: f64, a3: f64) -> f64 { |
| 79 |
use crate::polyeval::d_polyeval4; |
| 80 |
d_polyeval4(x, a0, a1, a2, a3) |
| 81 |
} |
| 82 |
#[inline(always)] |
| 83 |
fn halley(&self, x: f64, a: f64) -> f64 { |
| 84 |
use crate::cube_roots::cbrtf::halley_refine_d_fma; |
| 85 |
halley_refine_d_fma(x, a) |
| 86 |
} |
| 87 |
#[inline(always)] |
| 88 |
fn exact_mul(&self, x: f64, y: f64) -> DoubleDouble { |
| 89 |
DoubleDouble::from_exact_mult_fma(x, y) |
| 90 |
} |
| 91 |
#[inline(always)] |
| 92 |
fn quick_mult_f64(&self, x: DoubleDouble, y: f64) -> DoubleDouble { |
| 93 |
DoubleDouble::quick_mult_f64_fma(x, y) |
| 94 |
} |
| 95 |
} |
| 96 |
|
| 97 |
#[inline(always)] |
| 98 |
fn cbrt_gen_impl<B: CbrtBackend>(x: f64, backend: B) -> f64 { |
| 99 |
// 1; 2^{1/3}; 2^{2/3} |
| 100 |
static ESCALE: [f64; 3] = [ |
| 101 |
1.0, |
| 102 |
f64::from_bits(0x3ff428a2f98d728b), |
| 103 |
f64::from_bits(0x3ff965fea53d6e3d), |
| 104 |
]; |
| 105 |
|
| 106 |
let bits = x.to_bits(); |
| 107 |
let mut exp = ((bits >> 52) & 0x7ff) as i32; |
| 108 |
let mut mant = bits & ((1u64 << 52) - 1); |
| 109 |
|
| 110 |
if exp == 0x7ff || x == 0.0 { |
| 111 |
return x + x; |
| 112 |
} |
| 113 |
|
| 114 |
// Normalize subnormal |
| 115 |
if exp == 0 && x != 0.0 { |
| 116 |
let norm = x * f64::from_bits(0x4350000000000000); // * 2^54 |
| 117 |
let norm_bits = norm.to_bits(); |
| 118 |
mant = norm_bits & ((1u64 << 52) - 1); |
| 119 |
exp = ((norm_bits >> 52) & 0x7ff) as i32 - 54; |
| 120 |
} |
| 121 |
|
| 122 |
exp -= 1023; |
| 123 |
|
| 124 |
mant |= 0x3ff << 52; |
| 125 |
let m = f64::from_bits(mant); |
| 126 |
|
| 127 |
// Polynomial for x^(1/3) on [1.0; 2.0] |
| 128 |
// Generated by Sollya: |
| 129 |
// d = [1.0, 2.0]; |
| 130 |
// f_cbrt = x^(1/3); |
| 131 |
// Q = fpminimax(f_cbrt, 4, [|D...|], d, relative, floating); |
| 132 |
// See ./notes/cbrt.sollya |
| 133 |
|
| 134 |
let p = backend.polyeval4( |
| 135 |
m, |
| 136 |
f64::from_bits(0x3fe1b0babceeaafa), |
| 137 |
f64::from_bits(0x3fe2c9a3e8e06a3c), |
| 138 |
f64::from_bits(0xbfc4dc30afb71885), |
| 139 |
f64::from_bits(0x3f97a8d3e05458e4), |
| 140 |
); |
| 141 |
|
| 142 |
// split exponent e = 3*q + r with r in {0,1,2} |
| 143 |
// use div_euclid/rem_euclid to get r >= 0 |
| 144 |
let q = exp.div_euclid(3); |
| 145 |
let rem_scale = exp.rem_euclid(3); |
| 146 |
|
| 147 |
let z = p * ESCALE[rem_scale as usize]; |
| 148 |
|
| 149 |
let mm = fast_ldexp(m, rem_scale); // bring mantissa into [1;8] |
| 150 |
|
| 151 |
let r = 1.0 / mm; |
| 152 |
|
| 153 |
// One Halley's method step |
| 154 |
// then refine in partial double-double precision with Newton-Raphson iteration |
| 155 |
let y0 = backend.halley(z, mm); |
| 156 |
let d2y = backend.exact_mul(y0, y0); |
| 157 |
let d3y = backend.quick_mult_f64(d2y, y0); |
| 158 |
// Newton-Raphson step |
| 159 |
// h = (x^3 - a) * r |
| 160 |
// y1 = y0 - 1/3 * h * y0 |
| 161 |
let h = ((d3y.hi - mm) + d3y.lo) * r; |
| 162 |
// y1 = y0 - 1/3*y0*(h.lo + h.hi) = y0 - 1/3 *y0*h.lo - 1/3 * y0 * h.hi |
| 163 |
let y = backend.fma(-f64::from_bits(0x3fd5555555555555), y0 * h, y0); |
| 164 |
|
| 165 |
f64::copysign(fast_ldexp(y, q), x) |
| 166 |
} |
| 167 |
|
| 168 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 169 |
#[target_feature(enable = "avx", enable = "fma")] |
| 170 |
unsafe fn cbrt_fma_impl(x: f64) -> f64 { |
| 171 |
cbrt_gen_impl(x, FmaCbrtBackend {}) |
| 172 |
} |
| 173 |
|
| 174 |
/// Computes cube root |
| 175 |
/// |
| 176 |
/// Max found ULP 0.5 |
| 177 |
pub fn f_cbrt(x: f64) -> f64 { |
| 178 |
#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] |
| 179 |
{ |
| 180 |
cbrt_gen_impl(x, GenericCbrtBackend {}) |
| 181 |
} |
| 182 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 183 |
{ |
| 184 |
use std::sync::OnceLock; |
| 185 |
static EXECUTOR: OnceLock<unsafe fn(f64) -> f64> = OnceLock::new(); |
| 186 |
let q = EXECUTOR.get_or_init(|| { |
| 187 |
if std::arch::is_x86_feature_detected!("avx") |
| 188 |
&& std::arch::is_x86_feature_detected!("fma") |
| 189 |
{ |
| 190 |
cbrt_fma_impl |
| 191 |
} else { |
| 192 |
fn def_cbrt(x: f64) -> f64 { |
| 193 |
cbrt_gen_impl(x, GenericCbrtBackend {}) |
| 194 |
} |
| 195 |
def_cbrt |
| 196 |
} |
| 197 |
}); |
| 198 |
unsafe { q(x) } |
| 199 |
} |
| 200 |
} |
| 201 |
|
| 202 |
#[cfg(test)] |
| 203 |
mod tests { |
| 204 |
use super::*; |
| 205 |
|
| 206 |
#[test] |
| 207 |
fn test_cbrt() { |
| 208 |
assert_eq!(f_cbrt(0.000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000005432309223745), |
| 209 |
0.000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000017579026781511548); |
| 210 |
assert_eq!(f_cbrt(1.225158611559834), 1.0700336588124544); |
| 211 |
assert_eq!(f_cbrt(0.00000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000139491540182158), 1.1173329935611586e-103); |
| 212 |
assert_eq!(f_cbrt(27.0), 3.0); |
| 213 |
assert_eq!(f_cbrt(64.0), 4.0); |
| 214 |
assert_eq!(f_cbrt(125.0), 5.0); |
| 215 |
assert_eq!(f_cbrt(216.0), 6.0); |
| 216 |
assert_eq!(f_cbrt(343.0), 7.0); |
| 217 |
assert_eq!(f_cbrt(512.0), 8.0); |
| 218 |
assert_eq!(f_cbrt(729.0), 9.0); |
| 219 |
assert_eq!(f_cbrt(-729.0), -9.0); |
| 220 |
assert_eq!(f_cbrt(-512.0), -8.0); |
| 221 |
assert_eq!(f_cbrt(-343.0), -7.0); |
| 222 |
assert_eq!(f_cbrt(-216.0), -6.0); |
| 223 |
assert_eq!(f_cbrt(-125.0), -5.0); |
| 224 |
assert_eq!(f_cbrt(-64.0), -4.0); |
| 225 |
assert_eq!(f_cbrt(-27.0), -3.0); |
| 226 |
assert_eq!(f_cbrt(0.0), 0.0); |
| 227 |
assert_eq!(f_cbrt(f64::INFINITY), f64::INFINITY); |
| 228 |
assert_eq!(f_cbrt(f64::NEG_INFINITY), f64::NEG_INFINITY); |
| 229 |
assert!(f_cbrt(f64::NAN).is_nan()); |
| 230 |
} |
| 231 |
} |