| 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::rounding::CpuRoundTiesEven; |
| 31 |
use std::hint::black_box; |
| 32 |
|
| 33 |
static TB: [u64; 32] = [ |
| 34 |
0x3fe0000000000000, |
| 35 |
0x3fe059b0d3158574, |
| 36 |
0x3fe0b5586cf9890f, |
| 37 |
0x3fe11301d0125b51, |
| 38 |
0x3fe172b83c7d517b, |
| 39 |
0x3fe1d4873168b9aa, |
| 40 |
0x3fe2387a6e756238, |
| 41 |
0x3fe29e9df51fdee1, |
| 42 |
0x3fe306fe0a31b715, |
| 43 |
0x3fe371a7373aa9cb, |
| 44 |
0x3fe3dea64c123422, |
| 45 |
0x3fe44e086061892d, |
| 46 |
0x3fe4bfdad5362a27, |
| 47 |
0x3fe5342b569d4f82, |
| 48 |
0x3fe5ab07dd485429, |
| 49 |
0x3fe6247eb03a5585, |
| 50 |
0x3fe6a09e667f3bcd, |
| 51 |
0x3fe71f75e8ec5f74, |
| 52 |
0x3fe7a11473eb0187, |
| 53 |
0x3fe82589994cce13, |
| 54 |
0x3fe8ace5422aa0db, |
| 55 |
0x3fe93737b0cdc5e5, |
| 56 |
0x3fe9c49182a3f090, |
| 57 |
0x3fea5503b23e255d, |
| 58 |
0x3feae89f995ad3ad, |
| 59 |
0x3feb7f76f2fb5e47, |
| 60 |
0x3fec199bdd85529c, |
| 61 |
0x3fecb720dcef9069, |
| 62 |
0x3fed5818dcfba487, |
| 63 |
0x3fedfc97337b9b5f, |
| 64 |
0x3feea4afa2a490da, |
| 65 |
0x3fef50765b6e4540, |
| 66 |
]; |
| 67 |
|
| 68 |
#[cold] |
| 69 |
fn coshf_accurate(z: f64, ia: f64, sp: u64, sm: u64) -> f32 { |
| 70 |
const CH: [u64; 7] = [ |
| 71 |
0x3ff0000000000000, |
| 72 |
0x3f962e42fefa39ef, |
| 73 |
0x3f2ebfbdff82c58f, |
| 74 |
0x3ebc6b08d702e0ed, |
| 75 |
0x3e43b2ab6fb92e5e, |
| 76 |
0x3dc5d886e6d54203, |
| 77 |
0x3d4430976b8ce6ef, |
| 78 |
]; |
| 79 |
|
| 80 |
const ILN2H: f64 = f64::from_bits(0x4047154765000000); |
| 81 |
const ILN2L: f64 = f64::from_bits(0x3e55c17f0bbbe880); |
| 82 |
let h = f_fmla(ILN2L, z, f_fmla(ILN2H, z, -ia)); |
| 83 |
let h2 = h * h; |
| 84 |
|
| 85 |
let q0 = f_fmla(h2, f64::from_bits(CH[6]), f64::from_bits(CH[4])); |
| 86 |
let q1 = f_fmla(h2, f64::from_bits(CH[2]), f64::from_bits(CH[0])); |
| 87 |
|
| 88 |
let te = f_fmla(h2 * h2, q0, q1); |
| 89 |
|
| 90 |
let j0 = f_fmla(h2, f64::from_bits(CH[5]), f64::from_bits(CH[3])); |
| 91 |
|
| 92 |
let to = f_fmla(h2, j0, f64::from_bits(CH[1])); |
| 93 |
|
| 94 |
let z0 = f_fmla(h, to, te); |
| 95 |
let z1 = f_fmla(-h, to, te); |
| 96 |
f_fmla(f64::from_bits(sp), z0, f64::from_bits(sm) * z1) as f32 |
| 97 |
} |
| 98 |
|
| 99 |
/// Hyperbolic cos |
| 100 |
/// |
| 101 |
/// Max found ULP 0.4999996 |
| 102 |
#[inline] |
| 103 |
pub fn f_coshf(x: f32) -> f32 { |
| 104 |
const C: [u64; 4] = [ |
| 105 |
0x3ff0000000000000, |
| 106 |
0x3f962e42fef4c4e7, |
| 107 |
0x3f2ebfd1b232f475, |
| 108 |
0x3ebc6b19384ecd93, |
| 109 |
]; |
| 110 |
|
| 111 |
const I_LN2: f64 = f64::from_bits(0x40471547652b82fe); |
| 112 |
let t = x.to_bits(); |
| 113 |
let z = x as f64; |
| 114 |
let ax = t.wrapping_shl(1); |
| 115 |
if ax > 0x8565a9f8u32 { |
| 116 |
// |x| >~ 89.4 |
| 117 |
if ax >= 0xff000000u32 { |
| 118 |
if ax.wrapping_shl(8) != 0 { |
| 119 |
return x + x; |
| 120 |
} // nan |
| 121 |
return f32::INFINITY; // +-inf |
| 122 |
} |
| 123 |
let r = black_box(2.0) * black_box(f64::from_bits(0x47efffffe0000000) as f32); |
| 124 |
return r; |
| 125 |
} |
| 126 |
if ax < 0x7c000000u32 { |
| 127 |
// |x| < 0.125 |
| 128 |
if ax < 0x74000000u32 { |
| 129 |
// |x| < 0.000488281 |
| 130 |
if ax < 0x66000000u32 { |
| 131 |
// |x| < 5.96046e-08 |
| 132 |
#[cfg(any( |
| 133 |
all( |
| 134 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 135 |
target_feature = "fma" |
| 136 |
), |
| 137 |
target_arch = "aarch64" |
| 138 |
))] |
| 139 |
{ |
| 140 |
use crate::common::f_fmlaf; |
| 141 |
return f_fmlaf(x.abs(), f64::from_bits(0x3e60000000000000) as f32, 1.0); |
| 142 |
} |
| 143 |
#[cfg(not(any( |
| 144 |
all( |
| 145 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 146 |
target_feature = "fma" |
| 147 |
), |
| 148 |
target_arch = "aarch64" |
| 149 |
)))] |
| 150 |
{ |
| 151 |
let dx = x as f64; |
| 152 |
return f_fmla(dx.abs(), f64::from_bits(0x3e60000000000000), 1.0) as f32; |
| 153 |
} |
| 154 |
} |
| 155 |
#[cfg(any( |
| 156 |
all( |
| 157 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 158 |
target_feature = "fma" |
| 159 |
), |
| 160 |
target_arch = "aarch64" |
| 161 |
))] |
| 162 |
{ |
| 163 |
use crate::common::f_fmlaf; |
| 164 |
return f_fmlaf(0.5 * x, x, 1.0); |
| 165 |
} |
| 166 |
#[cfg(not(any( |
| 167 |
all( |
| 168 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 169 |
target_feature = "fma" |
| 170 |
), |
| 171 |
target_arch = "aarch64" |
| 172 |
)))] |
| 173 |
{ |
| 174 |
let dx = x as f64; |
| 175 |
return f_fmla(0.5 * dx, dx, 1.0) as f32; |
| 176 |
} |
| 177 |
} |
| 178 |
const CP: [u64; 4] = [ |
| 179 |
0x3fdfffffffffffe3, |
| 180 |
0x3fa55555555723cf, |
| 181 |
0x3f56c16bee4a5986, |
| 182 |
0x3efa0483fc0328f7, |
| 183 |
]; |
| 184 |
let z2 = z * z; |
| 185 |
let z4 = z2 * z2; |
| 186 |
let w0 = f_fmla(z2, f64::from_bits(CP[1]), f64::from_bits(CP[0])); |
| 187 |
let w1 = f_fmla(z2, f64::from_bits(CP[3]), f64::from_bits(CP[2])); |
| 188 |
return f_fmla(z2, f_fmla(z4, w1, w0), 1.0) as f32; |
| 189 |
} |
| 190 |
let a = I_LN2 * z; |
| 191 |
let ia = a.cpu_round_ties_even(); |
| 192 |
let h = a - ia; |
| 193 |
let h2 = h * h; |
| 194 |
let ja = (ia + f64::from_bits(0x4338000000000000)).to_bits(); |
| 195 |
let jp: i64 = ja as i64; |
| 196 |
let jm = -jp; |
| 197 |
let sp = TB[(jp & 31) as usize].wrapping_add(jp.wrapping_shr(5).wrapping_shl(52) as u64); |
| 198 |
let sm = TB[(jm & 31) as usize].wrapping_add(jm.wrapping_shr(5).wrapping_shl(52) as u64); |
| 199 |
let te = f_fmla(h2, f64::from_bits(C[2]), f64::from_bits(C[0])); |
| 200 |
let to = f_fmla(h2, f64::from_bits(C[3]), f64::from_bits(C[1])); |
| 201 |
let rp = f64::from_bits(sp) * f_fmla(h, to, te); |
| 202 |
let rm = f64::from_bits(sm) * f_fmla(-h, to, te); |
| 203 |
let r = rp + rm; |
| 204 |
let ub = r; |
| 205 |
let lb = r - f64::from_bits(0x3de3edbbe4560327) * r; |
| 206 |
// Ziv's accuracy test |
| 207 |
if ub != lb { |
| 208 |
return coshf_accurate(z, ia, sp, sm); |
| 209 |
} |
| 210 |
ub as f32 |
| 211 |
} |
| 212 |
|
| 213 |
#[cfg(test)] |
| 214 |
mod tests { |
| 215 |
use super::*; |
| 216 |
|
| 217 |
#[test] |
| 218 |
fn test_coshf() { |
| 219 |
assert_eq!(f_coshf(-0.5), 1.127626); |
| 220 |
assert_eq!(f_coshf(0.5), 1.127626); |
| 221 |
assert_eq!(f_coshf(7.), 548.317); |
| 222 |
} |
| 223 |
} |