| 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::exponents::expf::{ExpfBackend, GenericExpfBackend}; |
| 30 |
|
| 31 |
#[inline(always)] |
| 32 |
fn expm1f_gen<B: ExpfBackend>(x: f32, backend: B) -> f32 { |
| 33 |
let x_u: u32 = x.to_bits(); |
| 34 |
let x_abs = x_u & 0x7fff_ffffu32; |
| 35 |
|
| 36 |
// When |x| > 25*log(2), or nan |
| 37 |
if x_abs >= 0x418a_a123u32 { |
| 38 |
// x < log(2^-25) |
| 39 |
if x.is_sign_negative() { |
| 40 |
// exp(-Inf) = 0 |
| 41 |
if x.is_infinite() { |
| 42 |
return -1.0; |
| 43 |
} |
| 44 |
// exp(nan) = nan |
| 45 |
if x.is_nan() { |
| 46 |
return x; |
| 47 |
} |
| 48 |
return -1.0; |
| 49 |
} else { |
| 50 |
// x >= 89 or nan |
| 51 |
if x_u >= 0x42b2_0000 { |
| 52 |
return x + f32::INFINITY; |
| 53 |
} |
| 54 |
} |
| 55 |
} |
| 56 |
|
| 57 |
// |x| < 2^-4 |
| 58 |
if x_abs < 0x3d80_0000u32 { |
| 59 |
// |x| < 2^-25 |
| 60 |
if x_abs < 0x3300_0000u32 { |
| 61 |
// x = -0.0f |
| 62 |
if x_u == 0x8000_0000u32 { |
| 63 |
return x; |
| 64 |
} |
| 65 |
// When |x| < 2^-25, the relative error of the approximation e^x - 1 ~ x |
| 66 |
// is: |
| 67 |
// |(e^x - 1) - x| / |e^x - 1| < |x^2| / |x| |
| 68 |
// = |x| |
| 69 |
// < 2^-25 |
| 70 |
// < epsilon(1)/2. |
| 71 |
// To simplify the rounding decision and make it more efficient, we use |
| 72 |
// fma(x, x, x) ~ x + x^2 instead. |
| 73 |
// Note: to use the formula x + x^2 to decide the correct rounding, we |
| 74 |
// do need fma(x, x, x) to prevent underflow caused by x*x when |x| < |
| 75 |
// 2^-76. For targets without FMA instructions, we simply use double for |
| 76 |
// intermediate results as it is more efficient than using an emulated |
| 77 |
// version of FMA. |
| 78 |
#[cfg(any( |
| 79 |
all( |
| 80 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 81 |
target_feature = "fma" |
| 82 |
), |
| 83 |
target_arch = "aarch64" |
| 84 |
))] |
| 85 |
{ |
| 86 |
return backend.fmaf(x, x, x); |
| 87 |
} |
| 88 |
#[cfg(not(any( |
| 89 |
all( |
| 90 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 91 |
target_feature = "fma" |
| 92 |
), |
| 93 |
target_arch = "aarch64" |
| 94 |
)))] |
| 95 |
{ |
| 96 |
let xd = x as f64; |
| 97 |
return backend.fma(xd, xd, xd) as f32; |
| 98 |
} |
| 99 |
} |
| 100 |
|
| 101 |
const C: [u64; 7] = [ |
| 102 |
0x3fe0000000000000, |
| 103 |
0x3fc55555555557dd, |
| 104 |
0x3fa55555555552fa, |
| 105 |
0x3f8111110fcd58b7, |
| 106 |
0x3f56c16c1717660b, |
| 107 |
0x3f2a0241f0006d62, |
| 108 |
0x3efa01e3f8d3c060, |
| 109 |
]; |
| 110 |
|
| 111 |
// 2^-25 <= |x| < 2^-4 |
| 112 |
let xd = x as f64; |
| 113 |
let xsq = xd * xd; |
| 114 |
// Degree-8 minimax polynomial generated by Sollya with: |
| 115 |
// > display = hexadecimal; |
| 116 |
// > P = fpminimax((expm1(x) - x)/x^2, 6, [|D...|], [-2^-4, 2^-4]); |
| 117 |
|
| 118 |
return backend.fma( |
| 119 |
backend.polyeval7( |
| 120 |
xd, |
| 121 |
f64::from_bits(C[0]), |
| 122 |
f64::from_bits(C[1]), |
| 123 |
f64::from_bits(C[2]), |
| 124 |
f64::from_bits(C[3]), |
| 125 |
f64::from_bits(C[4]), |
| 126 |
f64::from_bits(C[5]), |
| 127 |
f64::from_bits(C[6]), |
| 128 |
), |
| 129 |
xsq, |
| 130 |
xd, |
| 131 |
) as f32; |
| 132 |
} |
| 133 |
|
| 134 |
// For -104 < x < 89, to compute expm1(x), we perform the following range |
| 135 |
// reduction: find hi, mid, lo such that: |
| 136 |
// x = hi + mid + lo, in which |
| 137 |
// hi is an integer, |
| 138 |
// mid * 2^7 is an integer |
| 139 |
// -2^(-8) <= lo < 2^-8. |
| 140 |
// In particular, |
| 141 |
// hi + mid = round(x * 2^7) * 2^(-7). |
| 142 |
// Then, |
| 143 |
// expm1(x) = expm1(hi + mid + lo) = exp(hi) * exp(mid) * exp(lo) - 1. |
| 144 |
// We store exp(hi) and exp(mid) in the lookup tables EXP_M1 and EXP_M2 |
| 145 |
// respectively. exp(lo) is computed using a degree-4 minimax polynomial |
| 146 |
// generated by Sollya. |
| 147 |
|
| 148 |
// x_hi = (hi + mid) * 2^7 = round(x * 2^7). |
| 149 |
let kf = backend.roundf(x * 128.); |
| 150 |
// Subtract (hi + mid) from x to get lo. |
| 151 |
let xd = backend.fmaf(kf, -0.0078125 /* - 1/128 */, x) as f64; |
| 152 |
let mut x_hi = unsafe { kf.to_int_unchecked::<i32>() }; // it's already not indeterminate. |
| 153 |
x_hi += 104 << 7; |
| 154 |
// hi = x_hi >> 7 |
| 155 |
let exp_hi = f64::from_bits(crate::exponents::expf::EXP_M1[(x_hi >> 7) as usize]); |
| 156 |
// mid * 2^7 = x_hi & 0x0000'007fU; |
| 157 |
let exp_mid = f64::from_bits(crate::exponents::expf::EXP_M2[(x_hi & 0x7f) as usize]); |
| 158 |
// Degree-4 minimax polynomial generated by Sollya with the following |
| 159 |
// commands: |
| 160 |
// d = [-2^-8, 2^-8]; |
| 161 |
// f_exp = expm1(x)/x; |
| 162 |
// Q = fpminimax(f_exp, 3, [|D...|], [-2^-8, 2^-8]); |
| 163 |
let p = backend.polyeval5( |
| 164 |
xd, |
| 165 |
1., |
| 166 |
f64::from_bits(0x3feffffffffff777), |
| 167 |
f64::from_bits(0x3fe000000000071c), |
| 168 |
f64::from_bits(0x3fc555566668e5e7), |
| 169 |
f64::from_bits(0x3fa55555555ef243), |
| 170 |
); |
| 171 |
backend.fma(p * exp_hi, exp_mid, -1.) as f32 |
| 172 |
} |
| 173 |
|
| 174 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 175 |
#[target_feature(enable = "avx", enable = "fma")] |
| 176 |
unsafe fn expm1f_fma_impl(x: f32) -> f32 { |
| 177 |
use crate::exponents::expf::FmaBackend; |
| 178 |
expm1f_gen(x, FmaBackend {}) |
| 179 |
} |
| 180 |
|
| 181 |
/// Computes e^x - 1 |
| 182 |
/// |
| 183 |
/// Max ULP 0.5 |
| 184 |
#[inline] |
| 185 |
pub fn f_expm1f(x: f32) -> f32 { |
| 186 |
#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] |
| 187 |
{ |
| 188 |
expm1f_gen(x, GenericExpfBackend {}) |
| 189 |
} |
| 190 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 191 |
{ |
| 192 |
use std::sync::OnceLock; |
| 193 |
static EXECUTOR: OnceLock<unsafe fn(f32) -> f32> = OnceLock::new(); |
| 194 |
let q = EXECUTOR.get_or_init(|| { |
| 195 |
if std::arch::is_x86_feature_detected!("avx") |
| 196 |
&& std::arch::is_x86_feature_detected!("fma") |
| 197 |
{ |
| 198 |
expm1f_fma_impl |
| 199 |
} else { |
| 200 |
fn def_expm1f(x: f32) -> f32 { |
| 201 |
expm1f_gen(x, GenericExpfBackend {}) |
| 202 |
} |
| 203 |
def_expm1f |
| 204 |
} |
| 205 |
}); |
| 206 |
unsafe { q(x) } |
| 207 |
} |
| 208 |
} |
| 209 |
|
| 210 |
#[cfg(test)] |
| 211 |
mod tests { |
| 212 |
use crate::f_expm1f; |
| 213 |
|
| 214 |
#[test] |
| 215 |
fn test_expm1f() { |
| 216 |
assert_eq!(f_expm1f(-3.0923562e-5), -3.0923085e-5); |
| 217 |
assert_eq!(f_expm1f(2.213121), 8.144211); |
| 218 |
assert_eq!(f_expm1f(-3.213121), -0.9597691); |
| 219 |
assert_eq!(f_expm1f(-2.35099e-38), -2.35099e-38); |
| 220 |
assert_eq!( |
| 221 |
f_expm1f(0.00000000000000000000000000000000000004355616), |
| 222 |
0.00000000000000000000000000000000000004355616 |
| 223 |
); |
| 224 |
assert_eq!(f_expm1f(25.12315), 81441420000.0); |
| 225 |
assert_eq!(f_expm1f(12.986543), 436498.6); |
| 226 |
assert_eq!(f_expm1f(-12.986543), -0.99999774); |
| 227 |
assert_eq!(f_expm1f(-25.12315), -1.0); |
| 228 |
assert_eq!(f_expm1f(f32::INFINITY), f32::INFINITY); |
| 229 |
assert_eq!(f_expm1f(f32::NEG_INFINITY), -1.); |
| 230 |
assert!(f_expm1f(f32::NAN).is_nan()); |
| 231 |
} |
| 232 |
} |