| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 8/2025. 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::*; |
| 30 |
use crate::compound::compound_m1f::compoundf_expf_poly; |
| 31 |
use crate::compound::compoundf::{ |
| 32 |
COMPOUNDF_EXP2_T, COMPOUNDF_EXP2_U, LOG2P1_COMPOUNDF_INV, LOG2P1_COMPOUNDF_LOG2_INV, |
| 33 |
}; |
| 34 |
use crate::rounding::CpuRoundTiesEven; |
| 35 |
use std::hint::black_box; |
| 36 |
|
| 37 |
#[inline] |
| 38 |
fn powm1f_log2_fast(x: f64) -> f64 { |
| 39 |
/* for x > 0, 1+x is exact when 2^-29 <= x < 2^53 |
| 40 |
for x < 0, 1+x is exact when -1 < x <= 2^-30 */ |
| 41 |
|
| 42 |
// double u = (x >= 0x1p53) ? x : 1.0 + x; |
| 43 |
/* For x < 0x1p53, x + 1 is exact thus u = x+1. |
| 44 |
For x >= 2^53, we estimate log2(x) instead of log2(1+x), |
| 45 |
since log2(1+x) = log2(x) + log2(1+1/x), |
| 46 |
log2(x) >= 53 and |log2(1+1/x)| < 2^-52.471, the additional relative |
| 47 |
error is bounded by 2^-52.471/53 < 2^-58.198 */ |
| 48 |
|
| 49 |
let mut v = x.to_bits(); |
| 50 |
let m: u64 = v & 0xfffffffffffffu64; |
| 51 |
let e: i64 = (v >> 52) as i64 - 0x3ff + (m >= 0x6a09e667f3bcdu64) as i64; |
| 52 |
// 2^e/sqrt(2) < u < 2^e*sqrt(2), with -29 <= e <= 128 |
| 53 |
v = v.wrapping_sub((e << 52) as u64); |
| 54 |
let t = f64::from_bits(v); |
| 55 |
// u = 2^e*t with 1/sqrt(2) < t < sqrt(2) |
| 56 |
// thus log2(u) = e + log2(t) |
| 57 |
v = (f64::from_bits(v) + 2.0).to_bits(); // add 2 so that v.f is always in the binade [2, 4) |
| 58 |
let i = (v >> 45) as i32 - 0x2002d; // 0 <= i <= 45 |
| 59 |
let r = f64::from_bits(LOG2P1_COMPOUNDF_INV[i as usize]); |
| 60 |
let z = dd_fmla(r, t, -1.0); // exact, -1/64 <= z <= 1/64 |
| 61 |
// we approximates log2(t) by -log2(r) + log2(r*t) |
| 62 |
let p = crate::compound::compoundf::log2p1_polyeval_1(z); |
| 63 |
// p approximates log2(r*t) with rel. error < 2^-49.642, and |p| < 2^-5.459 |
| 64 |
e as f64 + (f64::from_bits(LOG2P1_COMPOUNDF_LOG2_INV[i as usize].1) + p) |
| 65 |
} |
| 66 |
|
| 67 |
/// Computes x^y - 1 |
| 68 |
pub fn f_powm1f(x: f32, y: f32) -> f32 { |
| 69 |
let ax: u32 = x.to_bits().wrapping_shl(1); |
| 70 |
let ay: u32 = y.to_bits().wrapping_shl(1); |
| 71 |
|
| 72 |
// filter out exceptional cases |
| 73 |
if ax == 0 || ax >= 0xffu32 << 24 || ay == 0 || ay >= 0xffu32 << 24 { |
| 74 |
if x.is_nan() || y.is_nan() { |
| 75 |
return f32::NAN; |
| 76 |
} |
| 77 |
|
| 78 |
// Handle infinities |
| 79 |
if x.is_infinite() { |
| 80 |
return if x.is_sign_positive() { |
| 81 |
if y.is_infinite() { |
| 82 |
return f32::INFINITY; |
| 83 |
} else if y > 0.0 { |
| 84 |
f32::INFINITY // inf^positive -> inf |
| 85 |
} else if y < 0.0 { |
| 86 |
-1.0 // inf^negative -> 0, so powm1 = -1 |
| 87 |
} else { |
| 88 |
f32::NAN // inf^0 -> NaN (0^0 conventionally 1, inf^0 = NaN) |
| 89 |
} |
| 90 |
} else { |
| 91 |
// x = -inf |
| 92 |
if y.is_infinite() { |
| 93 |
return -1.0; |
| 94 |
} |
| 95 |
if is_integerf(y) { |
| 96 |
// Negative base: (-inf)^even = +inf, (-inf)^odd = -inf |
| 97 |
let pow = if y as i32 % 2 == 0 { |
| 98 |
f32::INFINITY |
| 99 |
} else { |
| 100 |
f32::NEG_INFINITY |
| 101 |
}; |
| 102 |
pow - 1.0 |
| 103 |
} else { |
| 104 |
f32::NAN // Negative base with non-integer exponent |
| 105 |
} |
| 106 |
}; |
| 107 |
} |
| 108 |
|
| 109 |
// Handle y infinite |
| 110 |
if y.is_infinite() { |
| 111 |
return if x.abs() > 1.0 { |
| 112 |
if y.is_sign_positive() { |
| 113 |
f32::INFINITY |
| 114 |
} else { |
| 115 |
-1.0 |
| 116 |
} |
| 117 |
} else if x.abs() < 1.0 { |
| 118 |
if y.is_sign_positive() { |
| 119 |
-1.0 |
| 120 |
} else { |
| 121 |
f32::INFINITY |
| 122 |
} |
| 123 |
} else { |
| 124 |
// |x| == 1 |
| 125 |
f32::NAN // 1^inf or -1^inf is undefined |
| 126 |
}; |
| 127 |
} |
| 128 |
|
| 129 |
// Handle zero base |
| 130 |
if x == 0.0 { |
| 131 |
return if y > 0.0 { |
| 132 |
-1.0 // 0^positive -> 0, powm1 = -1 |
| 133 |
} else if y < 0.0 { |
| 134 |
f32::INFINITY // 0^negative -> inf |
| 135 |
} else { |
| 136 |
0.0 // 0^0 -> conventionally 1, powm1 = 0 |
| 137 |
}; |
| 138 |
} |
| 139 |
} |
| 140 |
|
| 141 |
let y_integer = is_integerf(y); |
| 142 |
|
| 143 |
let mut negative_parity: bool = false; |
| 144 |
|
| 145 |
let mut x = x; |
| 146 |
|
| 147 |
// Handle negative base with non-integer exponent |
| 148 |
if x < 0.0 { |
| 149 |
if !y_integer { |
| 150 |
return f32::NAN; // x < 0 and non-integer y |
| 151 |
} |
| 152 |
x = x.abs(); |
| 153 |
if is_odd_integerf(y) { |
| 154 |
negative_parity = true; |
| 155 |
} |
| 156 |
} |
| 157 |
|
| 158 |
let xd = x as f64; |
| 159 |
let yd = y as f64; |
| 160 |
let tx = xd.to_bits(); |
| 161 |
let ty = yd.to_bits(); |
| 162 |
|
| 163 |
let l: f64 = powm1f_log2_fast(f64::from_bits(tx)); |
| 164 |
|
| 165 |
/* l approximates log2(1+x) with relative error < 2^-47.997, |
| 166 |
and 2^-149 <= |l| < 128 */ |
| 167 |
|
| 168 |
let dt = l * f64::from_bits(ty); |
| 169 |
let t: u64 = dt.to_bits(); |
| 170 |
|
| 171 |
// detect overflow/underflow |
| 172 |
if (t.wrapping_shl(1)) >= (0x406u64 << 53) { |
| 173 |
// |t| >= 128 |
| 174 |
if t >= 0x3018bu64 << 46 { |
| 175 |
// t <= -150 |
| 176 |
return -1.; |
| 177 |
} else if (t >> 63) == 0 { |
| 178 |
// t >= 128: overflow |
| 179 |
return black_box(f32::from_bits(0x7e800000)) * black_box(f32::from_bits(0x7e800000)); |
| 180 |
} |
| 181 |
} |
| 182 |
|
| 183 |
let res = powm1_exp2m1_fast(f64::from_bits(t)); |
| 184 |
// For x < 0 and integer y = n: |
| 185 |
// if n is even: x^n = |x|^n → powm1 = |x|^n - 1 (same sign as res). |
| 186 |
// if n is odd: x^n = -|x|^n → powm1 = -|x|^n - 1 = - (|x|^n + 1). |
| 187 |
if negative_parity { |
| 188 |
(-res - 2.) as f32 |
| 189 |
} else { |
| 190 |
res as f32 |
| 191 |
} |
| 192 |
} |
| 193 |
|
| 194 |
#[inline] |
| 195 |
pub(crate) fn powm1_exp2m1_fast(t: f64) -> f64 { |
| 196 |
let k = t.cpu_round_ties_even(); // 0 <= |k| <= 150 |
| 197 |
let mut r = t - k; // |r| <= 1/2, exact |
| 198 |
let mut v: f64 = 3.015625 + r; // 2.5 <= v <= 3.5015625 |
| 199 |
// we add 2^-6 so that i is rounded to nearest |
| 200 |
let i: i32 = (v.to_bits() >> 46) as i32 - 0x10010; // 0 <= i <= 32 |
| 201 |
r -= f64::from_bits(COMPOUNDF_EXP2_T[i as usize]); // exact |
| 202 |
// now |r| <= 2^-6 |
| 203 |
// 2^t = 2^k * exp2_U[i][0] * 2^r |
| 204 |
let mut s = f64::from_bits(COMPOUNDF_EXP2_U[i as usize].1); |
| 205 |
let su = unsafe { |
| 206 |
k.to_int_unchecked::<i64>().wrapping_shl(52) // k is already integer |
| 207 |
}; |
| 208 |
s = f64::from_bits(s.to_bits().wrapping_add(su as u64)); |
| 209 |
let q_poly = compoundf_expf_poly(r); |
| 210 |
v = q_poly; |
| 211 |
|
| 212 |
#[cfg(any( |
| 213 |
all( |
| 214 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 215 |
target_feature = "fma" |
| 216 |
), |
| 217 |
target_arch = "aarch64" |
| 218 |
))] |
| 219 |
{ |
| 220 |
v = f_fmla(v, s, s - 1f64); |
| 221 |
} |
| 222 |
#[cfg(not(any( |
| 223 |
all( |
| 224 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 225 |
target_feature = "fma" |
| 226 |
), |
| 227 |
target_arch = "aarch64" |
| 228 |
)))] |
| 229 |
{ |
| 230 |
use crate::double_double::DoubleDouble; |
| 231 |
let p0 = DoubleDouble::from_full_exact_add(s, -1.); |
| 232 |
let z = DoubleDouble::from_exact_mult(v, s); |
| 233 |
v = DoubleDouble::add(z, p0).to_f64(); |
| 234 |
} |
| 235 |
|
| 236 |
v |
| 237 |
} |
| 238 |
|
| 239 |
#[cfg(test)] |
| 240 |
mod tests { |
| 241 |
use super::*; |
| 242 |
#[test] |
| 243 |
fn test_powm1f() { |
| 244 |
assert_eq!(f_powm1f(1.83329e-40, 2.4645883e-32), -2.2550315e-30); |
| 245 |
assert_eq!(f_powm1f(f32::INFINITY, f32::INFINITY), f32::INFINITY); |
| 246 |
assert_eq!(f_powm1f(-3.375, -9671689000000000000000000.), -1.); |
| 247 |
assert_eq!(f_powm1f(3., 2.), 8.); |
| 248 |
assert_eq!(f_powm1f(3., 3.), 26.); |
| 249 |
assert_eq!(f_powm1f(5., 2.), 24.); |
| 250 |
assert_eq!(f_powm1f(5., -2.), 1. / 25. - 1.); |
| 251 |
assert_eq!(f_powm1f(-5., 2.), 24.); |
| 252 |
assert_eq!(f_powm1f(-5., 3.), -126.); |
| 253 |
assert_eq!( |
| 254 |
f_powm1f(196560., 0.000000000000000000000000000000000000001193773), |
| 255 |
1.455057e-38 |
| 256 |
); |
| 257 |
assert!(f_powm1f(f32::NAN, f32::INFINITY).is_nan()); |
| 258 |
assert!(f_powm1f(f32::INFINITY, f32::NAN).is_nan()); |
| 259 |
} |
| 260 |
} |