| 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 |
|
| 31 |
/// Hyperbolic tangent |
| 32 |
/// |
| 33 |
/// Max found ULP 0.4999994 |
| 34 |
#[inline] |
| 35 |
pub fn f_tanhf(x: f32) -> f32 { |
| 36 |
let z = x as f64; |
| 37 |
let t = x.to_bits(); |
| 38 |
let ux = t; |
| 39 |
let e = ux.wrapping_shr(23) & 0xff; |
| 40 |
if e == 0xff { |
| 41 |
if ux << 9 != 0 { |
| 42 |
return x + x; |
| 43 |
} // x = nan |
| 44 |
const IR: [f32; 2] = [1.0, -1.0]; |
| 45 |
return IR[ux.wrapping_shr(31) as usize]; // x = +-inf |
| 46 |
} |
| 47 |
if e < 115 { |
| 48 |
// |x| < 2^-13 |
| 49 |
if e < 102 { |
| 50 |
// |x| < 2^-26 |
| 51 |
if ux.wrapping_shl(1) == 0 { |
| 52 |
return x; |
| 53 |
} |
| 54 |
#[cfg(any( |
| 55 |
all( |
| 56 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 57 |
target_feature = "fma" |
| 58 |
), |
| 59 |
target_arch = "aarch64" |
| 60 |
))] |
| 61 |
{ |
| 62 |
use crate::common::f_fmlaf; |
| 63 |
let res = f_fmlaf(-x, x.abs(), x); |
| 64 |
return res; |
| 65 |
} |
| 66 |
#[cfg(not(any( |
| 67 |
all( |
| 68 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 69 |
target_feature = "fma" |
| 70 |
), |
| 71 |
target_arch = "aarch64" |
| 72 |
)))] |
| 73 |
{ |
| 74 |
let dx = x as f64; |
| 75 |
let res = f_fmla(-dx, dx.abs(), dx); |
| 76 |
return res as f32; |
| 77 |
} |
| 78 |
} |
| 79 |
#[cfg(any( |
| 80 |
all( |
| 81 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 82 |
target_feature = "fma" |
| 83 |
), |
| 84 |
target_arch = "aarch64" |
| 85 |
))] |
| 86 |
{ |
| 87 |
use crate::common::f_fmlaf; |
| 88 |
let x2 = x * x; |
| 89 |
return f_fmlaf(x, -f64::from_bits(0x3fd5555560000000) as f32 * x2, x); |
| 90 |
} |
| 91 |
#[cfg(not(any( |
| 92 |
all( |
| 93 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 94 |
target_feature = "fma" |
| 95 |
), |
| 96 |
target_arch = "aarch64" |
| 97 |
)))] |
| 98 |
{ |
| 99 |
let dx = x as f64; |
| 100 |
let x2 = dx * dx; |
| 101 |
return f_fmla(dx, -f64::from_bits(0x3fd5555560000000) * x2, dx) as f32; |
| 102 |
} |
| 103 |
} |
| 104 |
if ux.wrapping_shl(1) > (0x41102cb3u32 << 1) { |
| 105 |
return f32::copysign(1.0, x) - f32::copysign(f64::from_bits(0x3e60000000000000) as f32, x); |
| 106 |
} |
| 107 |
let z2 = z * z; |
| 108 |
let z4 = z2 * z2; |
| 109 |
let z8 = z4 * z4; |
| 110 |
const CN: [u64; 8] = [ |
| 111 |
0x3ff0000000000000, |
| 112 |
0x3fc30877b8b72d33, |
| 113 |
0x3f7694aa09ae9e5e, |
| 114 |
0x3f14101377abb729, |
| 115 |
0x3e9e0392b1db0018, |
| 116 |
0x3e12533756e546f7, |
| 117 |
0x3d6d62e5abe6ae8a, |
| 118 |
0x3c9b06be534182de, |
| 119 |
]; |
| 120 |
const CD: [u64; 8] = [ |
| 121 |
0x3ff0000000000000, |
| 122 |
0x3fded99131b0ebea, |
| 123 |
0x3fa0d27ed6c95a69, |
| 124 |
0x3f47cbdaca0e9fcc, |
| 125 |
0x3edb4e60b892578e, |
| 126 |
0x3e5a6f707c5c71ab, |
| 127 |
0x3dc35a8b6e2cd94c, |
| 128 |
0x3d0ca8230677aa01, |
| 129 |
]; |
| 130 |
let mut n0 = f_fmla(z2, f64::from_bits(CN[1]), f64::from_bits(CN[0])); |
| 131 |
let n2 = f_fmla(z2, f64::from_bits(CN[3]), f64::from_bits(CN[2])); |
| 132 |
let mut n4 = f_fmla(z2, f64::from_bits(CN[5]), f64::from_bits(CN[4])); |
| 133 |
let n6 = f_fmla(z2, f64::from_bits(CN[7]), f64::from_bits(CN[6])); |
| 134 |
n0 = f_fmla(z4, n2, n0); |
| 135 |
n4 = f_fmla(z4, n6, n4); |
| 136 |
n0 = f_fmla(z8, n4, n0); |
| 137 |
let mut d0 = f_fmla(z2, f64::from_bits(CD[1]), f64::from_bits(CD[0])); |
| 138 |
let d2 = f_fmla(z2, f64::from_bits(CD[3]), f64::from_bits(CD[2])); |
| 139 |
let mut d4 = f_fmla(z2, f64::from_bits(CD[5]), f64::from_bits(CD[4])); |
| 140 |
let d6 = f_fmla(z2, f64::from_bits(CD[7]), f64::from_bits(CD[6])); |
| 141 |
d0 = f_fmla(z4, d2, d0); |
| 142 |
d4 = f_fmla(z4, d6, d4); |
| 143 |
d0 = f_fmla(z8, d4, d0); |
| 144 |
let r = z * n0 / d0; |
| 145 |
r as f32 |
| 146 |
} |
| 147 |
|
| 148 |
#[cfg(test)] |
| 149 |
mod tests { |
| 150 |
use super::*; |
| 151 |
|
| 152 |
#[test] |
| 153 |
fn test_tanhf() { |
| 154 |
assert_eq!(f_tanhf(-0.5), -0.46211717); |
| 155 |
assert_eq!(f_tanhf(0.5), 0.46211717); |
| 156 |
assert_eq!(f_tanhf(7.), 0.99999833); |
| 157 |
} |
| 158 |
} |