| 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 std::hint::black_box; |
| 31 |
|
| 32 |
#[inline(always)] |
| 33 |
pub(crate) fn poly12<Q: Fn(f64, f64, f64) -> f64>(z: f64, c: [u64; 12], fma: &Q) -> f64 { |
| 34 |
let z2 = z * z; |
| 35 |
let z4 = z2 * z2; |
| 36 |
let mut c0 = fma(z, f64::from_bits(c[1]), f64::from_bits(c[0])); |
| 37 |
let c2 = fma(z, f64::from_bits(c[3]), f64::from_bits(c[2])); |
| 38 |
let mut c4 = fma(z, f64::from_bits(c[5]), f64::from_bits(c[4])); |
| 39 |
let c6 = fma(z, f64::from_bits(c[7]), f64::from_bits(c[6])); |
| 40 |
let mut c8 = fma(z, f64::from_bits(c[9]), f64::from_bits(c[8])); |
| 41 |
let c10 = fma(z, f64::from_bits(c[11]), f64::from_bits(c[10])); |
| 42 |
c0 = fma(c2, z2, c0); |
| 43 |
c4 = fma(c6, z2, c4); |
| 44 |
c8 = fma(z2, c10, c8); |
| 45 |
fma(z4, fma(z4, c8, c4), c0) |
| 46 |
} |
| 47 |
|
| 48 |
#[cold] |
| 49 |
fn as_special(x: f32) -> f32 { |
| 50 |
const PIH: f32 = f64::from_bits(0x400921fb60000000) as f32; |
| 51 |
const PIL: f32 = -f64::from_bits(0x3e70000000000000) as f32; |
| 52 |
let t = x.to_bits(); |
| 53 |
if t == (0x7fu32 << 23) { |
| 54 |
return 0.0; |
| 55 |
} // x=1 |
| 56 |
if t == (0x17fu32 << 23) { |
| 57 |
return PIH + PIL; |
| 58 |
} // x=-1 |
| 59 |
let ax = t.wrapping_shl(1); |
| 60 |
if ax > (0xffu32 << 24) { |
| 61 |
return x + x; |
| 62 |
} // nan |
| 63 |
f32::NAN |
| 64 |
} |
| 65 |
|
| 66 |
#[inline(always)] |
| 67 |
fn acosf_gen_impl<Q: Fn(f64, f64, f64) -> f64>(x: f32, fma: Q) -> f32 { |
| 68 |
const PI2: f64 = f64::from_bits(0x3ff921fb54442d18); |
| 69 |
const O: [f64; 2] = [0., f64::from_bits(0x400921fb54442d18)]; |
| 70 |
let xs = x as f64; |
| 71 |
let mut r: f64; |
| 72 |
let t = x.to_bits(); |
| 73 |
let ax = t.wrapping_shl(1); |
| 74 |
if ax >= 0x7f << 24 { |
| 75 |
return as_special(x); |
| 76 |
} |
| 77 |
if ax < 0x7ec2a1dcu32 { |
| 78 |
// |x| < 0.880141 |
| 79 |
const B: [u64; 16] = [ |
| 80 |
0x3fefffffffd9ccb8, |
| 81 |
0x3fc5555c94838007, |
| 82 |
0x3fb32ded4b7c20fa, |
| 83 |
0x3fa8566df703309e, |
| 84 |
0xbf9980c959bec9a3, |
| 85 |
0x3fe56fbb04998344, |
| 86 |
0xc01403d8e4c49f52, |
| 87 |
0x403b06c3e9f311ea, |
| 88 |
0xc059ea97c4e2c21f, |
| 89 |
0x407200b8261cc61b, |
| 90 |
0xc082274c2799a5c7, |
| 91 |
0x408a558a59cc19d3, |
| 92 |
0xc08aca4b6a529ff0, |
| 93 |
0x408228744703f813, |
| 94 |
0xc06d7dbb0b322228, |
| 95 |
0x4045c2018c0c0105, |
| 96 |
]; |
| 97 |
/* avoid spurious underflow */ |
| 98 |
if ax < 0x40000000u32 { |
| 99 |
// |x| < 2^-63 |
| 100 |
return PI2 as f32; |
| 101 |
} |
| 102 |
let z = xs; |
| 103 |
let z2 = z * z; |
| 104 |
|
| 105 |
let w0 = fma(z2, f64::from_bits(B[1]), f64::from_bits(B[0])); |
| 106 |
let w1 = fma(z2, f64::from_bits(B[3]), f64::from_bits(B[2])); |
| 107 |
let w2 = fma(z2, f64::from_bits(B[5]), f64::from_bits(B[4])); |
| 108 |
let w3 = fma(z2, f64::from_bits(B[7]), f64::from_bits(B[6])); |
| 109 |
let w4 = fma(z2, f64::from_bits(B[9]), f64::from_bits(B[8])); |
| 110 |
let w5 = fma(z2, f64::from_bits(B[11]), f64::from_bits(B[10])); |
| 111 |
let w6 = fma(z2, f64::from_bits(B[13]), f64::from_bits(B[12])); |
| 112 |
let w7 = fma(z2, f64::from_bits(B[15]), f64::from_bits(B[14])); |
| 113 |
|
| 114 |
let z4 = z2 * z2; |
| 115 |
let z8 = z4 * z4; |
| 116 |
let z16 = z8 * z8; |
| 117 |
|
| 118 |
r = z |
| 119 |
* ((fma(z4, w1, w0) + z8 * fma(z4, w3, w2)) |
| 120 |
+ z16 * (fma(z4, w5, w4) + z8 * fma(z4, w7, w6))); |
| 121 |
|
| 122 |
let ub = f64::from_bits(0x3ff921fb54574191) - r; |
| 123 |
let lb = f64::from_bits(0x3ff921fb543118a0) - r; |
| 124 |
// Ziv's accuracy test |
| 125 |
if ub == lb { |
| 126 |
return ub as f32; |
| 127 |
} |
| 128 |
} |
| 129 |
// accurate path |
| 130 |
if ax < (0x7eu32 << 24) { |
| 131 |
const C: [u64; 12] = [ |
| 132 |
0x3fc555555555529c, |
| 133 |
0x3fb333333337e0dd, |
| 134 |
0x3fa6db6db3b4465e, |
| 135 |
0x3f9f1c72e13ac306, |
| 136 |
0x3f96e89cebe06bc4, |
| 137 |
0x3f91c6dcf5289094, |
| 138 |
0x3f8c6dbbcc7c6315, |
| 139 |
0x3f88f8dc2615e996, |
| 140 |
0x3f7a5833b7bf15e8, |
| 141 |
0x3f943f44ace1665c, |
| 142 |
0xbf90fb17df881c73, |
| 143 |
0x3fa07520c026b2d6, |
| 144 |
]; |
| 145 |
if t == 0x328885a3u32 { |
| 146 |
return black_box(f64::from_bits(0x3ff921fb60000000) as f32) |
| 147 |
+ black_box(f64::from_bits(0x3e60000000000000) as f32); |
| 148 |
} |
| 149 |
if t == 0x39826222u32 { |
| 150 |
return black_box(f64::from_bits(0x3ff920f6a0000000) as f32) |
| 151 |
+ black_box(f64::from_bits(0x3e60000000000000) as f32); |
| 152 |
} |
| 153 |
let x2 = xs * xs; |
| 154 |
r = fma(-(xs * x2), poly12(x2, C, &fma), PI2 - xs); |
| 155 |
} else { |
| 156 |
const C: [u64; 12] = [ |
| 157 |
0x3ff6a09e667f3bcb, |
| 158 |
0x3fbe2b7dddff2db9, |
| 159 |
0x3f9b27247ab42dbc, |
| 160 |
0x3f802995cc4e0744, |
| 161 |
0x3f65ffb0276ec8ea, |
| 162 |
0x3f5033885a928dec, |
| 163 |
0x3f3911f2be23f8c7, |
| 164 |
0x3f24c3c55d2437fd, |
| 165 |
0x3f0af477e1d7b461, |
| 166 |
0x3f0abd6bdff67dcb, |
| 167 |
0xbef1717e86d0fa28, |
| 168 |
0x3ef6ff526de46023, |
| 169 |
]; |
| 170 |
let bx = xs.abs(); |
| 171 |
let z = 1.0 - bx; |
| 172 |
let s = f64::copysign(z.sqrt(), xs); |
| 173 |
r = fma(s, poly12(z, C, &fma), O[t.wrapping_shr(31) as usize]); |
| 174 |
} |
| 175 |
r as f32 |
| 176 |
} |
| 177 |
|
| 178 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 179 |
#[target_feature(enable = "avx", enable = "fma")] |
| 180 |
unsafe fn acosf_fma_impl(x: f32) -> f32 { |
| 181 |
acosf_gen_impl(x, f64::mul_add) |
| 182 |
} |
| 183 |
|
| 184 |
/// Compute acos |
| 185 |
/// |
| 186 |
/// Max found ULP 0.49999982 |
| 187 |
#[inline] |
| 188 |
pub fn f_acosf(x: f32) -> f32 { |
| 189 |
#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] |
| 190 |
{ |
| 191 |
acosf_gen_impl(x, f_fmla) |
| 192 |
} |
| 193 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 194 |
{ |
| 195 |
use std::sync::OnceLock; |
| 196 |
static EXECUTOR: OnceLock<unsafe fn(f32) -> f32> = OnceLock::new(); |
| 197 |
let q = EXECUTOR.get_or_init(|| { |
| 198 |
if std::arch::is_x86_feature_detected!("avx") |
| 199 |
&& std::arch::is_x86_feature_detected!("fma") |
| 200 |
{ |
| 201 |
acosf_fma_impl |
| 202 |
} else { |
| 203 |
fn def_acosf(x: f32) -> f32 { |
| 204 |
acosf_gen_impl(x, f_fmla) |
| 205 |
} |
| 206 |
def_acosf |
| 207 |
} |
| 208 |
}); |
| 209 |
unsafe { q(x) } |
| 210 |
} |
| 211 |
} |
| 212 |
|
| 213 |
#[cfg(test)] |
| 214 |
mod tests { |
| 215 |
use super::*; |
| 216 |
|
| 217 |
#[test] |
| 218 |
fn test_acosf() { |
| 219 |
assert_eq!(f_acosf(-0.5), 2.0943952); |
| 220 |
assert_eq!(f_acosf(0.5), std::f32::consts::FRAC_PI_3); |
| 221 |
assert!(f_acosf(7.).is_nan()); |
| 222 |
} |
| 223 |
} |