| 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::logs::{LOG_R_DD, LOG_RANGE_REDUCTION}; |
| 31 |
use crate::polyeval::{f_estrin_polyeval8, f_polyeval6}; |
| 32 |
|
| 33 |
#[inline] |
| 34 |
pub(crate) fn core_logf(x: f64) -> f64 { |
| 35 |
let x_u = x.to_bits(); |
| 36 |
|
| 37 |
const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; |
| 38 |
|
| 39 |
let mut x_e: i32 = -(E_BIAS as i32); |
| 40 |
|
| 41 |
// log2(x) = log2(2^x_e * x_m) |
| 42 |
// = x_e + log2(x_m) |
| 43 |
// Range reduction for log2(x_m): |
| 44 |
// For each x_m, we would like to find r such that: |
| 45 |
// -2^-8 <= r * x_m - 1 < 2^-7 |
| 46 |
let shifted = (x_u >> 45) as i32; |
| 47 |
let index = shifted & 0x7F; |
| 48 |
let r = f64::from_bits(LOG_RANGE_REDUCTION[index as usize]); |
| 49 |
|
| 50 |
// Add unbiased exponent. Add an extra 1 if the 8 leading fractional bits are |
| 51 |
// all 1's. |
| 52 |
x_e = x_e.wrapping_add(x_u.wrapping_add(1u64 << 45).wrapping_shr(52) as i32); |
| 53 |
let e_x = x_e as f64; |
| 54 |
|
| 55 |
const LOG_2_HI: f64 = f64::from_bits(0x3fe62e42fefa3800); |
| 56 |
const LOG_2_LO: f64 = f64::from_bits(0x3d2ef35793c76730); |
| 57 |
|
| 58 |
let log_r_dd = LOG_R_DD[index as usize]; |
| 59 |
|
| 60 |
// hi is exact |
| 61 |
let hi = f_fmla(e_x, LOG_2_HI, f64::from_bits(log_r_dd.1)); |
| 62 |
let lo = f_fmla(e_x, LOG_2_LO, f64::from_bits(log_r_dd.0)); |
| 63 |
|
| 64 |
// Set m = 1.mantissa. |
| 65 |
let x_m = (x_u & 0x000F_FFFF_FFFF_FFFFu64) | 0x3FF0_0000_0000_0000u64; |
| 66 |
let m = f64::from_bits(x_m); |
| 67 |
|
| 68 |
let u; |
| 69 |
#[cfg(any( |
| 70 |
all( |
| 71 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 72 |
target_feature = "fma" |
| 73 |
), |
| 74 |
target_arch = "aarch64" |
| 75 |
))] |
| 76 |
{ |
| 77 |
u = f_fmla(r, m, -1.0); // exact |
| 78 |
} |
| 79 |
#[cfg(not(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::logs::LOG_CD; |
| 88 |
let c_m = x_m & 0x3FFF_E000_0000_0000u64; |
| 89 |
let c = f64::from_bits(c_m); |
| 90 |
u = f_fmla(r, m - c, f64::from_bits(LOG_CD[index as usize])); // exact |
| 91 |
} |
| 92 |
|
| 93 |
let r1 = hi; |
| 94 |
// Polynomial for log(1+x)/x generated in Sollya: |
| 95 |
// d = [-2^-8, 2^-7]; |
| 96 |
// f_log = log(1 + x)/x; |
| 97 |
// Q = fpminimax(f_log, 5, [|D...|], d); |
| 98 |
// See ./notes/log1pf_core.sollya |
| 99 |
let p = f_polyeval6( |
| 100 |
u, |
| 101 |
f64::from_bits(0x3fefffffffffffff), |
| 102 |
f64::from_bits(0xbfdffffffffff3e6), |
| 103 |
f64::from_bits(0x3fd5555555626b74), |
| 104 |
f64::from_bits(0xbfd0000026aeecc8), |
| 105 |
f64::from_bits(0x3fc9999114d16c06), |
| 106 |
f64::from_bits(0xbfc51e433a85278a), |
| 107 |
); |
| 108 |
f_fmla(p, u, r1) + lo |
| 109 |
} |
| 110 |
|
| 111 |
/// Computes log(x+1) |
| 112 |
/// |
| 113 |
/// Max ULP 0.5 |
| 114 |
#[inline] |
| 115 |
pub fn f_log1pf(x: f32) -> f32 { |
| 116 |
let ux = x.to_bits().wrapping_shl(1); |
| 117 |
if ux >= 0xffu32 << 24 || ux == 0 { |
| 118 |
// |x| == 0, |x| == inf, x == NaN |
| 119 |
if ux == 0 { |
| 120 |
return x; |
| 121 |
} |
| 122 |
if x.is_infinite() { |
| 123 |
return if x.is_sign_positive() { |
| 124 |
f32::INFINITY |
| 125 |
} else { |
| 126 |
f32::NAN |
| 127 |
}; |
| 128 |
} |
| 129 |
return x + f32::NAN; |
| 130 |
} |
| 131 |
|
| 132 |
let xd = x as f64; |
| 133 |
let ax = x.to_bits() & 0x7fff_ffffu32; |
| 134 |
|
| 135 |
// Use log1p(x) = log(1 + x) for |x| > 2^-6; |
| 136 |
if ax > 0x3c80_0000u32 { |
| 137 |
if x == -1. { |
| 138 |
return f32::NEG_INFINITY; |
| 139 |
} |
| 140 |
let x1p = xd + 1.; |
| 141 |
if x1p <= 0. { |
| 142 |
if x1p == 0. { |
| 143 |
return f32::NEG_INFINITY; |
| 144 |
} |
| 145 |
return f32::NAN; |
| 146 |
} |
| 147 |
return core_logf(x1p) as f32; |
| 148 |
} |
| 149 |
|
| 150 |
// log(1+x) is expected to be used near zero |
| 151 |
// Polynomial generated by Sollya: |
| 152 |
// d = [-2^-6; 2^-6]; |
| 153 |
// f_log1pf = log(1+x)/x; |
| 154 |
// Q = fpminimax(f_log1pf, 7, [|0, D...|], d); |
| 155 |
// See ./notes/log1pf.sollya |
| 156 |
|
| 157 |
let p = f_estrin_polyeval8( |
| 158 |
xd, |
| 159 |
f64::from_bits(0x3ff0000000000000), |
| 160 |
f64::from_bits(0xbfe0000000000000), |
| 161 |
f64::from_bits(0x3fd5555555556aad), |
| 162 |
f64::from_bits(0xbfd000000000181a), |
| 163 |
f64::from_bits(0x3fc999998998124e), |
| 164 |
f64::from_bits(0xbfc55555452e2a2b), |
| 165 |
f64::from_bits(0x3fc24adb8cde4aa7), |
| 166 |
f64::from_bits(0xbfc0019db915ef6f), |
| 167 |
) * xd; |
| 168 |
p as f32 |
| 169 |
} |
| 170 |
|
| 171 |
#[inline] |
| 172 |
pub(crate) fn core_log1pf(x: f32) -> f64 { |
| 173 |
let xd = x as f64; |
| 174 |
let ax = x.to_bits() & 0x7fff_ffffu32; |
| 175 |
|
| 176 |
// Use log1p(x) = log(1 + x) for |x| > 2^-6; |
| 177 |
if ax > 0x3c80_0000u32 { |
| 178 |
let x1p = xd + 1.; |
| 179 |
return core_logf(x1p); |
| 180 |
} |
| 181 |
|
| 182 |
// log(1+x) is expected to be used near zero |
| 183 |
// Polynomial generated by Sollya: |
| 184 |
// d = [-2^-6; 2^-6]; |
| 185 |
// f_log1pf = log(1+x)/x; |
| 186 |
// Q = fpminimax(f_log1pf, 7, [|0, D...|], d); |
| 187 |
// See ./notes/log1pf.sollya |
| 188 |
|
| 189 |
f_estrin_polyeval8( |
| 190 |
xd, |
| 191 |
f64::from_bits(0x3ff0000000000000), |
| 192 |
f64::from_bits(0xbfe0000000000000), |
| 193 |
f64::from_bits(0x3fd5555555556aad), |
| 194 |
f64::from_bits(0xbfd000000000181a), |
| 195 |
f64::from_bits(0x3fc999998998124e), |
| 196 |
f64::from_bits(0xbfc55555452e2a2b), |
| 197 |
f64::from_bits(0x3fc24adb8cde4aa7), |
| 198 |
f64::from_bits(0xbfc0019db915ef6f), |
| 199 |
) * xd |
| 200 |
} |
| 201 |
|
| 202 |
#[cfg(test)] |
| 203 |
mod tests { |
| 204 |
use super::*; |
| 205 |
|
| 206 |
#[test] |
| 207 |
fn log1pf_works() { |
| 208 |
assert!(f_log1pf(f32::from_bits(0xffefb9a7)).is_nan()); |
| 209 |
assert!(f_log1pf(f32::NAN).is_nan()); |
| 210 |
assert_eq!(f_log1pf(f32::from_bits(0x41078feb)), 2.2484074); |
| 211 |
assert_eq!(f_log1pf(-0.0000014305108), -0.0000014305118); |
| 212 |
assert_eq!(f_log1pf(0.0), 0.0); |
| 213 |
assert_eq!(f_log1pf(2.0), 1.0986123); |
| 214 |
assert_eq!(f_log1pf(-0.7), -1.2039728); |
| 215 |
assert_eq!(f_log1pf(-0.0000000000043243), -4.3243e-12); |
| 216 |
assert_eq!(f_log1pf(f32::INFINITY), f32::INFINITY); |
| 217 |
assert!(f_log1pf(-2.0).is_nan()); |
| 218 |
assert!(f_log1pf(f32::NAN).is_nan()); |
| 219 |
} |
| 220 |
} |