| 1 |
/* |
| 2 |
* // Copyright (c) Radzivon Bartoshyk 9/2025. 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::{dyad_fmla, f_fmla}; |
| 30 |
use crate::double_double::DoubleDouble; |
| 31 |
use std::hint::black_box; |
| 32 |
|
| 33 |
// case hypot(x,y) >= 2^1024 |
| 34 |
#[cold] |
| 35 |
fn hypot_overflow() -> f64 { |
| 36 |
const Z: f64 = f64::from_bits(0x7fefffffffffffff); |
| 37 |
black_box(Z) + black_box(Z) |
| 38 |
} |
| 39 |
|
| 40 |
#[cold] |
| 41 |
fn hypot_denorm(a: u64, b: u64) -> f64 { |
| 42 |
let mut a = a; |
| 43 |
let mut b = b; |
| 44 |
let af = (a as i64) as f64; |
| 45 |
let bf = (b as i64) as f64; |
| 46 |
let mut underflow; |
| 47 |
// af and bf are x and y multiplied by 2^1074, thus integers |
| 48 |
a = a.wrapping_shl(1); |
| 49 |
b = b.wrapping_shl(1); |
| 50 |
let mut rm = f_fmla(af, af, bf * bf).sqrt() as u64; |
| 51 |
let mut tm: i64 = rm.wrapping_shl(1) as i64; |
| 52 |
let mut denom: i64 = a |
| 53 |
.wrapping_mul(a) |
| 54 |
.wrapping_add(b.wrapping_mul(b)) |
| 55 |
.wrapping_sub((tm as u64).wrapping_mul(tm as u64)) as i64; |
| 56 |
// D = a^2+b^2 - tm^2 |
| 57 |
while denom > 2 * tm { |
| 58 |
// tm too small |
| 59 |
denom -= 2 * tm + 1; // (tm+1)^2 = tm^2 + 2*tm + 1 |
| 60 |
tm = tm.wrapping_add(1); |
| 61 |
} |
| 62 |
while denom < 0 { |
| 63 |
// tm too large |
| 64 |
denom += 2 * tm - 1; // (tm-1)^2 = tm^2 - 2*tm + 1 |
| 65 |
tm = tm.wrapping_sub(1); |
| 66 |
} |
| 67 |
// tm = floor(sqrt(a^2+b^2)) and 0 <= D = a^2+b^2 - tm^2 < 2*tm+1 |
| 68 |
// if D=0 and tm is even, the result is exact |
| 69 |
// if D=0 and tm is odd, the result is a midpoint |
| 70 |
let rb: i32 = (tm & 1) as i32; // round bit for rm |
| 71 |
let rb2: i32 = (denom >= tm) as i32; // round bit for tm |
| 72 |
let sb: i32 = (denom != 0) as i32; // sticky bit for rm |
| 73 |
rm = (tm >> 1) as u64; // truncate the low bit |
| 74 |
underflow = rm < 0x10000000000000u64; |
| 75 |
if rb != 0 || sb != 0 { |
| 76 |
let op = 1.0 + f64::from_bits(0x3c90000000000000); |
| 77 |
let om = 1.0 - f64::from_bits(0x3c90000000000000); |
| 78 |
if op == om { |
| 79 |
// rounding to nearest |
| 80 |
if sb != 0 { |
| 81 |
rm = rm.wrapping_add(rb as u64); |
| 82 |
// we have no underflow when rm is now 2^52 and rb2 != 0 |
| 83 |
// Remark: we cannot have a^2+b^2 = (tm+1/2)^2 exactly |
| 84 |
// since this would mean a^2+b^2 = tm^2+tm+1/4, |
| 85 |
// thus a^2+b^2 would be an odd multiple of 2^-1077 |
| 86 |
// (since ulp(tm) = 2^-1075) |
| 87 |
if rm >> 52 != 0 && rb2 != 0 { |
| 88 |
underflow = false; |
| 89 |
} |
| 90 |
} else { |
| 91 |
// sticky bit is 0, round bit is 1: underflow doos not change |
| 92 |
rm += rm & 1; // even rounding |
| 93 |
} |
| 94 |
} else if op > 1.0 { |
| 95 |
// rounding upwards |
| 96 |
rm = rm.wrapping_add(1); |
| 97 |
// we have no underflow when rm is now 2^52 and tm was odd |
| 98 |
if (rm >> 52) != 0 && (tm & 1) != 0 { |
| 99 |
underflow = false; |
| 100 |
} |
| 101 |
} |
| 102 |
if underflow { |
| 103 |
// trigger underflow exception _after_ rounding for inexact results |
| 104 |
let trig_uf = black_box(f64::from_bits(0x0010000000000000)); |
| 105 |
_ = black_box(black_box(trig_uf) * black_box(trig_uf)); // triggers underflow |
| 106 |
} |
| 107 |
} |
| 108 |
// else the result is exact, and we have no underflow |
| 109 |
f64::from_bits(rm) |
| 110 |
} |
| 111 |
|
| 112 |
/* Here the square root is refined by Newton iterations: x^2+y^2 is exact |
| 113 |
and fits in a 128-bit integer, so the approximation is squared (which |
| 114 |
also fits in a 128-bit integer), compared and adjusted if necessary using |
| 115 |
the exact value of x^2+y^2. */ |
| 116 |
#[cold] |
| 117 |
fn hypot_hard(x: f64, y: f64) -> f64 { |
| 118 |
let op = 1.0 + f64::from_bits(0x3c90000000000000); |
| 119 |
let om = 1.0 - f64::from_bits(0x3c90000000000000); |
| 120 |
let mut xi = x.to_bits(); |
| 121 |
let yi = y.to_bits(); |
| 122 |
let mut bm = (xi & 0x000fffffffffffff) | (1u64 << 52); |
| 123 |
let mut lm = (yi & 0x000fffffffffffff) | (1u64 << 52); |
| 124 |
let be: i32 = (xi >> 52) as i32; |
| 125 |
let le: i32 = (yi >> 52) as i32; |
| 126 |
let ri = f_fmla(x, x, y * y).sqrt().to_bits(); |
| 127 |
const BS: u32 = 2; |
| 128 |
let mut rm: u64 = ri & 0x000fffffffffffff; |
| 129 |
let mut re: i32 = ((ri >> 52) as i32).wrapping_sub(0x3ff); |
| 130 |
rm |= 1u64 << 52; |
| 131 |
for _ in 0..3 { |
| 132 |
if rm == (1u64 << 52) { |
| 133 |
rm = 0x001fffffffffffff; |
| 134 |
re = re.wrapping_sub(1); |
| 135 |
} else { |
| 136 |
rm = rm.wrapping_sub(1); |
| 137 |
} |
| 138 |
} |
| 139 |
bm = bm.wrapping_shl(BS); |
| 140 |
let mut m2: u64 = bm.wrapping_mul(bm); |
| 141 |
let de: i32 = be.wrapping_sub(le); |
| 142 |
let mut ls: i32 = (BS as i32).wrapping_sub(de); |
| 143 |
if ls >= 0 { |
| 144 |
lm <<= ls; |
| 145 |
m2 = m2.wrapping_add(lm.wrapping_mul(lm)); |
| 146 |
} else { |
| 147 |
let lm2: u128 = (lm as u128).wrapping_mul(lm as u128); |
| 148 |
ls = ls.wrapping_mul(2); |
| 149 |
m2 = m2.wrapping_add((lm2 >> -ls) as u64); // since ls < 0, the shift by -ls is legitimate |
| 150 |
m2 |= ((lm2.wrapping_shl((128 + ls) as u32)) != 0) as u64; |
| 151 |
} |
| 152 |
let k: i32 = (BS as i32).wrapping_add(re); |
| 153 |
let mut denom: i64; |
| 154 |
loop { |
| 155 |
rm += 1 + (rm >= (1u64 << 53)) as u64; |
| 156 |
let tm: u64 = rm.wrapping_shl(k as u32); |
| 157 |
let rm2: u64 = tm.wrapping_mul(tm); |
| 158 |
denom = (m2 as i64).wrapping_sub(rm2 as i64); |
| 159 |
if denom <= 0 { |
| 160 |
break; |
| 161 |
} |
| 162 |
} |
| 163 |
if denom != 0 { |
| 164 |
if op == om { |
| 165 |
let ssm = if rm <= (1u64 << 53) { 1 } else { 0 }; |
| 166 |
let tm: u64 = (rm << k) - (1 << (k - ssm)); |
| 167 |
denom = (m2 as i64).wrapping_sub(tm.wrapping_mul(tm) as i64); |
| 168 |
if denom != 0 { |
| 169 |
rm = rm.wrapping_add((denom >> 63) as u64); |
| 170 |
} else { |
| 171 |
rm = rm.wrapping_sub(rm & 1); |
| 172 |
} |
| 173 |
} else { |
| 174 |
let ssm = if rm > (1u64 << 53) { 1u32 } else { 0 }; |
| 175 |
let pdm = if op == 1. { 1u64 } else { 0 }; |
| 176 |
rm = rm.wrapping_sub(pdm.wrapping_shl(ssm)); |
| 177 |
} |
| 178 |
} |
| 179 |
if rm >= (1u64 << 53) { |
| 180 |
rm >>= 1; |
| 181 |
re = re.wrapping_add(1); |
| 182 |
} |
| 183 |
|
| 184 |
let e: i64 = be.wrapping_sub(1).wrapping_add(re) as i64; |
| 185 |
xi = (e as u64).wrapping_shl(52).wrapping_add(rm); |
| 186 |
f64::from_bits(xi) |
| 187 |
} |
| 188 |
|
| 189 |
/// Computes hypot |
| 190 |
/// |
| 191 |
/// Max ULP 0.5 |
| 192 |
pub fn f_hypot(x: f64, y: f64) -> f64 { |
| 193 |
let xi = x.to_bits(); |
| 194 |
let yi = y.to_bits(); |
| 195 |
let emsk = 0x7ffu64 << 52; |
| 196 |
let mut ex = xi & emsk; |
| 197 |
let mut ey = yi & emsk; |
| 198 |
let mut x = x.abs(); |
| 199 |
let mut y = y.abs(); |
| 200 |
if ex == emsk || ey == emsk { |
| 201 |
/* Either x or y is NaN or Inf */ |
| 202 |
let wx = xi.wrapping_shl(1); |
| 203 |
let wy = yi.wrapping_shl(1); |
| 204 |
let wm = emsk.wrapping_shl(1); |
| 205 |
let ninf: i32 = ((wx == wm) ^ (wy == wm)) as i32; |
| 206 |
let nqnn: i32 = (((wx >> 52) == 0xfff) ^ ((wy >> 52) == 0xfff)) as i32; |
| 207 |
/* ninf is 1 when only one of x and y is +/-Inf |
| 208 |
nqnn is 1 when only one of x and y is qNaN |
| 209 |
IEEE 754 says that hypot(+/-Inf,qNaN)=hypot(qNaN,+/-Inf)=+Inf. */ |
| 210 |
if ninf != 0 && nqnn != 0 { |
| 211 |
return if wx == wm { x * x } else { y * y }; |
| 212 |
} |
| 213 |
return x + y; /* inf, nan */ |
| 214 |
} |
| 215 |
|
| 216 |
let u = f64::max(x, y); |
| 217 |
let v = f64::min(x, y); |
| 218 |
let mut xd = u.to_bits(); |
| 219 |
let mut yd = v.to_bits(); |
| 220 |
ey = yd; |
| 221 |
if (ey >> 52) == 0 { |
| 222 |
// y is subnormal |
| 223 |
if (yd) == 0 { |
| 224 |
return f64::from_bits(xd); |
| 225 |
} |
| 226 |
ex = xd; |
| 227 |
if (ex >> 52) == 0 { |
| 228 |
// x is subnormal too |
| 229 |
if ex == 0 { |
| 230 |
return 0.; |
| 231 |
} |
| 232 |
return hypot_denorm(ex, ey); |
| 233 |
} |
| 234 |
let nz = ey.leading_zeros(); |
| 235 |
ey = ey.wrapping_shl(nz - 11); |
| 236 |
ey &= u64::MAX >> 12; |
| 237 |
ey = ey.wrapping_sub(((nz as u64).wrapping_sub(12u64)) << 52); |
| 238 |
let t = ey; |
| 239 |
yd = t; |
| 240 |
} |
| 241 |
|
| 242 |
let de = xd.wrapping_sub(yd); |
| 243 |
if de > (27u64 << 52) { |
| 244 |
return dyad_fmla(f64::from_bits(0x3e40000000000000), v, u); |
| 245 |
} |
| 246 |
let off: i64 = ((0x3ffu64 << 52) as i64).wrapping_sub((xd & emsk) as i64); |
| 247 |
xd = xd.wrapping_add(off as u64); |
| 248 |
yd = yd.wrapping_add(off as u64); |
| 249 |
x = f64::from_bits(xd); |
| 250 |
y = f64::from_bits(yd); |
| 251 |
let x2 = DoubleDouble::from_exact_mult(x, x); |
| 252 |
let y2 = DoubleDouble::from_exact_mult(y, y); |
| 253 |
let r2 = x2.hi + y2.hi; |
| 254 |
let ir2 = 0.5 / r2; |
| 255 |
let dr2 = ((x2.hi - r2) + y2.hi) + (x2.lo + y2.lo); |
| 256 |
let mut th = r2.sqrt(); |
| 257 |
let rsqrt = DoubleDouble::from_exact_mult(th, ir2); |
| 258 |
let dz = dr2 - rsqrt.lo; |
| 259 |
let mut tl = rsqrt.hi * dz; |
| 260 |
let p = DoubleDouble::from_exact_add(th, tl); |
| 261 |
th = p.hi; |
| 262 |
tl = p.lo; |
| 263 |
let mut thd = th.to_bits(); |
| 264 |
let tld = tl.abs().to_bits(); |
| 265 |
ex = thd; |
| 266 |
ey = tld; |
| 267 |
ex &= 0x7ffu64 << 52; |
| 268 |
let aidr = ey.wrapping_add(0x3feu64 << 52).wrapping_sub(ex); |
| 269 |
let mid = (aidr.wrapping_sub(0x3c90000000000000).wrapping_add(16)) >> 5; |
| 270 |
if mid == 0 || aidr < 0x39b0000000000000u64 || aidr > 0x3c9fffffffffff80u64 { |
| 271 |
thd = hypot_hard(x, y).to_bits(); |
| 272 |
} |
| 273 |
thd = thd.wrapping_sub(off as u64); |
| 274 |
if thd >= (0x7ffu64 << 52) { |
| 275 |
return hypot_overflow(); |
| 276 |
} |
| 277 |
f64::from_bits(thd) |
| 278 |
} |
| 279 |
|
| 280 |
#[cfg(test)] |
| 281 |
mod tests { |
| 282 |
use super::*; |
| 283 |
#[test] |
| 284 |
fn test_hypot() { |
| 285 |
assert_eq!( |
| 286 |
f_hypot(2.8480945070593211E-306, 2.1219950320804403E-314), |
| 287 |
2.848094507059321e-306 |
| 288 |
); |
| 289 |
assert_eq!( |
| 290 |
f_hypot(3.5601181736115222E-307, 1.0609978954826362E-314), |
| 291 |
3.560118173611524e-307 |
| 292 |
); |
| 293 |
assert_eq!(f_hypot(2., 4.), 4.47213595499958); |
| 294 |
} |
| 295 |
} |