| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 7/2025. 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::dd_fmla; |
| 30 |
use crate::double_double::DoubleDouble; |
| 31 |
use crate::logs::log_dd::log_poly; |
| 32 |
use crate::logs::log_dd_coeffs::LOG_NEG_DD; |
| 33 |
use crate::polyeval::{f_estrin_polyeval7, f_polyeval3}; |
| 34 |
use crate::pow_tables::POW_INVERSE; |
| 35 |
|
| 36 |
#[inline(always)] |
| 37 |
pub(crate) fn log1p_tiny(z: f64) -> DoubleDouble { |
| 38 |
// See ./notes/log1p_tiny.sollya for generation |
| 39 |
const Q_1: [(u64, u64); 7] = [ |
| 40 |
(0xbc85555555555555, 0x3fd5555555555556), |
| 41 |
(0x0000000000000000, 0xbfd0000000000000), |
| 42 |
(0xbc6999999999999a, 0x3fc999999999999a), |
| 43 |
(0x3c75555555555555, 0xbfc5555555555556), |
| 44 |
(0x3c62492492492492, 0x3fc2492492492492), |
| 45 |
(0x0000000000000000, 0xbfc0000000000000), |
| 46 |
(0x3c5c71c71c71c71c, 0x3fbc71c71c71c71c), |
| 47 |
]; |
| 48 |
let mut r = DoubleDouble::quick_mul_f64_add_f64( |
| 49 |
DoubleDouble::from_bit_pair(Q_1[6]), |
| 50 |
z, |
| 51 |
f64::from_bits(0xbfc0000000000000), |
| 52 |
); |
| 53 |
r = DoubleDouble::quick_mul_f64_add(r, z, DoubleDouble::from_bit_pair(Q_1[4])); |
| 54 |
r = DoubleDouble::quick_mul_f64_add(r, z, DoubleDouble::from_bit_pair(Q_1[3])); |
| 55 |
r = DoubleDouble::quick_mul_f64_add(r, z, DoubleDouble::from_bit_pair(Q_1[2])); |
| 56 |
r = DoubleDouble::quick_mul_f64_add_f64(r, z, f64::from_bits(0xbfd0000000000000)); |
| 57 |
r = DoubleDouble::quick_mul_f64_add(r, z, DoubleDouble::from_bit_pair(Q_1[0])); |
| 58 |
r = DoubleDouble::quick_mul_f64_add_f64(r, z, f64::from_bits(0xbfe0000000000000)); |
| 59 |
r = DoubleDouble::quick_mul_f64_add_f64(r, z, f64::from_bits(0x3ff0000000000000)); |
| 60 |
DoubleDouble::quick_mult_f64(r, z) |
| 61 |
} |
| 62 |
|
| 63 |
#[inline(always)] |
| 64 |
fn log1p_poly_fast(z: f64) -> DoubleDouble { |
| 65 |
// Polynomial generated by Sollya |
| 66 |
// d = [-0.0040283203125,0.0040283203125]; |
| 67 |
// f = log(1+x); |
| 68 |
// p0 = x-x^2/2; |
| 69 |
// w = 1; |
| 70 |
// pf = p0+fpminimax(f-p0, [|3,4,5,6,7,8,9|], [|D...|], d, absolute, floating); |
| 71 |
// See ./notes/dd_log1p_fast.sollya |
| 72 |
const Q: [f64; 7] = [ |
| 73 |
f64::from_bits(0x3fd5555555555556), |
| 74 |
f64::from_bits(0xbfcfffffffffffdc), |
| 75 |
f64::from_bits(0x3fc99999998fd488), |
| 76 |
f64::from_bits(0xbfc5555555d90fc7), |
| 77 |
f64::from_bits(0x3fc24936d06d3bf8), |
| 78 |
f64::from_bits(0xbfbfff46726d8e88), |
| 79 |
f64::from_bits(0x3fa1847e2faea348), |
| 80 |
]; |
| 81 |
let x2 = DoubleDouble::from_exact_mult(z, z); |
| 82 |
let p = f_estrin_polyeval7(z, Q[0], Q[1], Q[2], Q[3], Q[4], Q[5], Q[6]); |
| 83 |
let mut t = DoubleDouble::quick_mult_f64(x2, p); |
| 84 |
t = DoubleDouble::quick_mult_f64(t, z); |
| 85 |
DoubleDouble::mul_f64_add(x2, -0.5, t) |
| 86 |
} |
| 87 |
|
| 88 |
#[inline(always)] |
| 89 |
fn log1p_tiny_fast(z: f64) -> DoubleDouble { |
| 90 |
// Polynomial generated by Sollya |
| 91 |
// d = [-2^-12,2^-12]; |
| 92 |
// f = log(1+x); |
| 93 |
// p0 = x-x^2/2; |
| 94 |
// w = 1; |
| 95 |
// pf = p0+fpminimax(f-p0, [|3,4,5,6|], [|107, D...|], d, absolute, floating); |
| 96 |
// See ./notes/dd_log1p_tiny_fast.sollya |
| 97 |
let x2 = DoubleDouble::from_exact_mult(z, z); |
| 98 |
let p = f_polyeval3( |
| 99 |
z, |
| 100 |
f64::from_bits(0xbfcffffffffffff4), |
| 101 |
f64::from_bits(0x3fc99999b4d2481f), |
| 102 |
f64::from_bits(0xbfc55555714a3cb8), |
| 103 |
); |
| 104 |
let DoubleDouble { hi: h, lo: r } = DoubleDouble::from_exact_mult(z, p); |
| 105 |
let DoubleDouble { hi: ph, lo: q } = DoubleDouble::add_f64( |
| 106 |
DoubleDouble::from_bit_pair((0xbc77e8068b994170, 0x3fd5555555555551)), |
| 107 |
h, |
| 108 |
); |
| 109 |
let p = DoubleDouble::new(r + q, ph); |
| 110 |
let mut t = DoubleDouble::quick_mult(x2, p); |
| 111 |
t = DoubleDouble::quick_mult_f64(t, z); |
| 112 |
let f = DoubleDouble::f64_add(z, DoubleDouble::mul_f64_add(x2, -0.5, t)); |
| 113 |
DoubleDouble::from_exact_add(f.hi, f.lo) |
| 114 |
} |
| 115 |
|
| 116 |
#[inline] |
| 117 |
pub(crate) fn log1p_dd(z: f64) -> DoubleDouble { |
| 118 |
let ax = z.to_bits().wrapping_shl(1); |
| 119 |
if ax < 0x7e60000000000000u64 { |
| 120 |
// |x| < 0x1p-12 |
| 121 |
return log1p_tiny(z); |
| 122 |
} |
| 123 |
let dz = DoubleDouble::from_full_exact_add(z, 1.0); |
| 124 |
|
| 125 |
// We'll compute log((z+1)+1) as log(xh+xl) = log(xh) + log(1+xl/xh). |
| 126 |
// since xl/xh < ulp(xh) we'll use for log(1+xl/xh) |
| 127 |
// one taylor term what means that log(1+xl/xh) = log_lo + O(x^2) |
| 128 |
|
| 129 |
let log_lo = if dz.hi <= f64::from_bits(0x7fd0000000000000) || dz.lo.abs() >= 4.0 { |
| 130 |
dz.lo / dz.hi |
| 131 |
} else { |
| 132 |
0. |
| 133 |
}; // avoid spurious underflow |
| 134 |
|
| 135 |
let x_u = dz.hi.to_bits(); |
| 136 |
let mut m = x_u & 0xfffffffffffff; |
| 137 |
let mut e: i64 = ((x_u >> 52) & 0x7ff) as i64; |
| 138 |
|
| 139 |
let t; |
| 140 |
if e != 0 { |
| 141 |
t = m | (0x3ffu64 << 52); |
| 142 |
m = m.wrapping_add(1u64 << 52); |
| 143 |
e -= 0x3ff; |
| 144 |
} else { |
| 145 |
/* x is a subnormal double */ |
| 146 |
let k = m.leading_zeros() - 11; |
| 147 |
|
| 148 |
e = -0x3fei64 - k as i64; |
| 149 |
m = m.wrapping_shl(k); |
| 150 |
t = m | (0x3ffu64 << 52); |
| 151 |
} |
| 152 |
|
| 153 |
/* now |x| = 2^_e*_t = 2^(_e-52)*m with 1 <= _t < 2, |
| 154 |
and 2^52 <= _m < 2^53 */ |
| 155 |
|
| 156 |
// log(x) = log(t) + E · log(2) |
| 157 |
let mut t = f64::from_bits(t); |
| 158 |
|
| 159 |
// If m > sqrt(2) we divide it by 2 so ensure 1/sqrt(2) < t < sqrt(2) |
| 160 |
let c: usize = (m >= 0x16a09e667f3bcd) as usize; |
| 161 |
static CY: [f64; 2] = [1.0, 0.5]; |
| 162 |
static CM: [u64; 2] = [44, 45]; |
| 163 |
|
| 164 |
e = e.wrapping_add(c as i64); |
| 165 |
let be = e; |
| 166 |
let i = m >> CM[c]; |
| 167 |
t *= CY[c]; |
| 168 |
|
| 169 |
let r = f64::from_bits(POW_INVERSE[(i - 181) as usize]); |
| 170 |
let log_r = DoubleDouble::from_bit_pair(LOG_NEG_DD[(i - 181) as usize]); |
| 171 |
|
| 172 |
let z = f64::mul_add(r, t, -1.0); |
| 173 |
|
| 174 |
const LOG2_DD: DoubleDouble = DoubleDouble::new( |
| 175 |
f64::from_bits(0x3c7abc9e3b39803f), |
| 176 |
f64::from_bits(0x3fe62e42fefa39ef), |
| 177 |
); |
| 178 |
|
| 179 |
let tt = DoubleDouble::mul_f64_add(LOG2_DD, be as f64, log_r); |
| 180 |
|
| 181 |
let v = DoubleDouble::full_add_f64(tt, z); |
| 182 |
let mut p = log_poly(z); |
| 183 |
// adding log(1+xl/xh) lower term |
| 184 |
p.lo += log_lo; |
| 185 |
DoubleDouble::f64_add(v.hi, DoubleDouble::new(v.lo + p.lo, p.hi)) |
| 186 |
} |
| 187 |
|
| 188 |
#[inline] |
| 189 |
pub(crate) fn log1p_fast_dd(z: f64) -> DoubleDouble { |
| 190 |
let ax = z.to_bits().wrapping_shl(1); |
| 191 |
if ax < 0x7e60000000000000u64 { |
| 192 |
// |x| < 2^-12 |
| 193 |
return log1p_tiny_fast(z); |
| 194 |
} |
| 195 |
let dz = DoubleDouble::from_full_exact_add(z, 1.0); |
| 196 |
|
| 197 |
// We'll compute log((z+1)+1) as log(xh+xl) = log(xh) + log(1+xl/xh). |
| 198 |
// since xl/xh < ulp(xh) we'll use for log(1+xl/xh) |
| 199 |
// one taylor term what means that log(1+xl/xh) = log_lo + O(x^2) |
| 200 |
|
| 201 |
let log_lo = if dz.hi <= f64::from_bits(0x7fd0000000000000) || dz.lo.abs() >= 4.0 { |
| 202 |
dz.lo / dz.hi |
| 203 |
} else { |
| 204 |
0. |
| 205 |
}; // avoid spurious underflow |
| 206 |
|
| 207 |
let x_u = dz.hi.to_bits(); |
| 208 |
let mut m = x_u & 0xfffffffffffff; |
| 209 |
let mut e: i64 = ((x_u >> 52) & 0x7ff) as i64; |
| 210 |
|
| 211 |
let t; |
| 212 |
if e != 0 { |
| 213 |
t = m | (0x3ffu64 << 52); |
| 214 |
m = m.wrapping_add(1u64 << 52); |
| 215 |
e -= 0x3ff; |
| 216 |
} else { |
| 217 |
/* x is a subnormal double */ |
| 218 |
let k = m.leading_zeros() - 11; |
| 219 |
|
| 220 |
e = -0x3fei64 - k as i64; |
| 221 |
m = m.wrapping_shl(k); |
| 222 |
t = m | (0x3ffu64 << 52); |
| 223 |
} |
| 224 |
|
| 225 |
/* now |x| = 2^_e*_t = 2^(_e-52)*m with 1 <= _t < 2, |
| 226 |
and 2^52 <= _m < 2^53 */ |
| 227 |
|
| 228 |
// log(x) = log(t) + E · log(2) |
| 229 |
let mut t = f64::from_bits(t); |
| 230 |
|
| 231 |
// If m > sqrt(2) we divide it by 2 so ensure 1/sqrt(2) < t < sqrt(2) |
| 232 |
let c: usize = (m >= 0x16a09e667f3bcd) as usize; |
| 233 |
static CY: [f64; 2] = [1.0, 0.5]; |
| 234 |
static CM: [u64; 2] = [44, 45]; |
| 235 |
|
| 236 |
e = e.wrapping_add(c as i64); |
| 237 |
let be = e; |
| 238 |
let i = m >> CM[c]; |
| 239 |
t *= CY[c]; |
| 240 |
|
| 241 |
let r = f64::from_bits(POW_INVERSE[(i - 181) as usize]); |
| 242 |
let log_r = DoubleDouble::from_bit_pair(LOG_NEG_DD[(i - 181) as usize]); |
| 243 |
|
| 244 |
let z = dd_fmla(r, t, -1.0); |
| 245 |
|
| 246 |
const LOG2_DD: DoubleDouble = DoubleDouble::new( |
| 247 |
f64::from_bits(0x3c7abc9e3b39803f), |
| 248 |
f64::from_bits(0x3fe62e42fefa39ef), |
| 249 |
); |
| 250 |
|
| 251 |
let tt = DoubleDouble::mul_f64_add(LOG2_DD, be as f64, log_r); |
| 252 |
|
| 253 |
let v = DoubleDouble::full_add_f64(tt, z); |
| 254 |
let mut p = log1p_poly_fast(z); |
| 255 |
// adding log(1+xl/xh) lower term |
| 256 |
p.lo += log_lo; |
| 257 |
DoubleDouble::f64_add(v.hi, DoubleDouble::new(v.lo + p.lo, p.hi)) |
| 258 |
} |