| 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::f_fmla; |
| 30 |
use crate::polyeval::f_polyeval6; |
| 31 |
use crate::sin_cosf::sincosf_eval::sincosf_eval; |
| 32 |
|
| 33 |
#[inline(always)] |
| 34 |
fn secf_gen_impl(x: f32) -> f32 { |
| 35 |
let x_abs = x.to_bits() & 0x7fff_ffffu32; |
| 36 |
let x = f32::from_bits(x_abs); |
| 37 |
let xd = x as f64; |
| 38 |
|
| 39 |
// |x| <= pi/16 |
| 40 |
if x_abs <= 0x3e49_0fdbu32 { |
| 41 |
// |x| < 0.000244141 |
| 42 |
if x_abs < 0x3980_0000u32 { |
| 43 |
// taylor series for sec(x) ~ 1 + x^2/2 + O(x^4) |
| 44 |
// for such small interval just doing 2 first coefficients from taylor series |
| 45 |
// FMA availability is mandatory to perform it in f32 without upcasting to f64. |
| 46 |
#[cfg(any( |
| 47 |
all( |
| 48 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 49 |
target_feature = "fma" |
| 50 |
), |
| 51 |
target_arch = "aarch64" |
| 52 |
))] |
| 53 |
{ |
| 54 |
use crate::common::f_fmlaf; |
| 55 |
return f_fmlaf(x, x * f32::from_bits(0x3f000000), 1.); |
| 56 |
} |
| 57 |
#[cfg(not(any( |
| 58 |
all( |
| 59 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 60 |
target_feature = "fma" |
| 61 |
), |
| 62 |
target_arch = "aarch64" |
| 63 |
)))] |
| 64 |
{ |
| 65 |
let x2 = xd * xd; |
| 66 |
return f_fmla(x2, f64::from_bits(0x3fe0000000000000), 1.) as f32; |
| 67 |
} |
| 68 |
} |
| 69 |
|
| 70 |
// Secant |
| 71 |
// Generated poly by Sollya: |
| 72 |
// f = 1 / cos(x); |
| 73 |
// d = [0.000244141; pi/16]; |
| 74 |
// pf = fpminimax(f, [|0, 2, 4, 6, 8, 10|], [|1, D...|], d, relative, floating); |
| 75 |
// |
| 76 |
// See ./notes/secf.sollya |
| 77 |
|
| 78 |
let x2 = xd * xd; |
| 79 |
let p = f_polyeval6( |
| 80 |
x2, |
| 81 |
f64::from_bits(0x3ff0000000000000), |
| 82 |
f64::from_bits(0x3fe000000001c0fb), |
| 83 |
f64::from_bits(0x3fcaaaaaa0b8a71b), |
| 84 |
f64::from_bits(0x3fb5b06437bc5a13), |
| 85 |
f64::from_bits(0x3fa192a33a9fca4f), |
| 86 |
f64::from_bits(0x3f8dde280c29af37), |
| 87 |
); |
| 88 |
return p as f32; |
| 89 |
} |
| 90 |
|
| 91 |
if x_abs >= 0x7f80_0000u32 { |
| 92 |
return x + f32::NAN; |
| 93 |
} |
| 94 |
|
| 95 |
// Formula: |
| 96 |
// cos(x) = cos((k + y)*pi/32) |
| 97 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 98 |
// The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
| 99 |
// and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
| 100 |
// computed using degree-7 and degree-6 minimax polynomials generated by |
| 101 |
// Sollya respectively. |
| 102 |
// Combine the results with the sine of sum formula: |
| 103 |
// cos(x) = cos((k + y)*pi/32) |
| 104 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 105 |
// = cosm1_y * cos_k + sin_y * sin_k |
| 106 |
// = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
| 107 |
// then sec(x) = 1/cos(x) |
| 108 |
|
| 109 |
let rs = sincosf_eval(xd, x_abs); |
| 110 |
(1. / f_fmla(rs.sin_y, -rs.sin_k, f_fmla(rs.cosm1_y, rs.cos_k, rs.cos_k))) as f32 |
| 111 |
} |
| 112 |
|
| 113 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 114 |
#[target_feature(enable = "avx", enable = "fma")] |
| 115 |
unsafe fn secf_fma_impl(x: f32) -> f32 { |
| 116 |
let x_abs = x.to_bits() & 0x7fff_ffffu32; |
| 117 |
let x = f32::from_bits(x_abs); |
| 118 |
let xd = x as f64; |
| 119 |
|
| 120 |
// |x| <= pi/16 |
| 121 |
if x_abs <= 0x3e49_0fdbu32 { |
| 122 |
// |x| < 0.000244141 |
| 123 |
if x_abs < 0x3980_0000u32 { |
| 124 |
// taylor series for sec(x) ~ 1 + x^2/2 + O(x^4) |
| 125 |
// for such small interval just doing 2 first coefficients from taylor series |
| 126 |
// FMA availability is mandatory to perform it in f32 without upcasting to f64. |
| 127 |
return f32::mul_add(x, x * f32::from_bits(0x3f000000), 1.); |
| 128 |
} |
| 129 |
|
| 130 |
// Secant |
| 131 |
// Generated poly by Sollya: |
| 132 |
// f = 1 / cos(x); |
| 133 |
// d = [0.000244141; pi/16]; |
| 134 |
// pf = fpminimax(f, [|0, 2, 4, 6, 8, 10|], [|1, D...|], d, relative, floating); |
| 135 |
// |
| 136 |
// See ./notes/secf.sollya |
| 137 |
|
| 138 |
let x2 = xd * xd; |
| 139 |
use crate::polyeval::d_polyeval6; |
| 140 |
let p = d_polyeval6( |
| 141 |
x2, |
| 142 |
f64::from_bits(0x3ff0000000000000), |
| 143 |
f64::from_bits(0x3fe000000001c0fb), |
| 144 |
f64::from_bits(0x3fcaaaaaa0b8a71b), |
| 145 |
f64::from_bits(0x3fb5b06437bc5a13), |
| 146 |
f64::from_bits(0x3fa192a33a9fca4f), |
| 147 |
f64::from_bits(0x3f8dde280c29af37), |
| 148 |
); |
| 149 |
return p as f32; |
| 150 |
} |
| 151 |
|
| 152 |
if x_abs >= 0x7f80_0000u32 { |
| 153 |
return x + f32::NAN; |
| 154 |
} |
| 155 |
|
| 156 |
// Formula: |
| 157 |
// cos(x) = cos((k + y)*pi/32) |
| 158 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 159 |
// The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
| 160 |
// and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
| 161 |
// computed using degree-7 and degree-6 minimax polynomials generated by |
| 162 |
// Sollya respectively. |
| 163 |
// Combine the results with the sine of sum formula: |
| 164 |
// cos(x) = cos((k + y)*pi/32) |
| 165 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 166 |
// = cosm1_y * cos_k + sin_y * sin_k |
| 167 |
// = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
| 168 |
// then sec(x) = 1/cos(x) |
| 169 |
use crate::sin_cosf::sincosf_eval::sincosf_eval_fma; |
| 170 |
let rs = sincosf_eval_fma(xd, x_abs); |
| 171 |
(1. / f64::mul_add( |
| 172 |
rs.sin_y, |
| 173 |
-rs.sin_k, |
| 174 |
f64::mul_add(rs.cosm1_y, rs.cos_k, rs.cos_k), |
| 175 |
)) as f32 |
| 176 |
} |
| 177 |
|
| 178 |
/// Computes secant ( 1 / cos(x) ) |
| 179 |
/// |
| 180 |
/// Max found ULP 0.5 |
| 181 |
#[inline] |
| 182 |
pub fn f_secf(x: f32) -> f32 { |
| 183 |
#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] |
| 184 |
{ |
| 185 |
secf_gen_impl(x) |
| 186 |
} |
| 187 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 188 |
{ |
| 189 |
use std::sync::OnceLock; |
| 190 |
static EXECUTOR: OnceLock<unsafe fn(f32) -> f32> = OnceLock::new(); |
| 191 |
let q = EXECUTOR.get_or_init(|| { |
| 192 |
if std::arch::is_x86_feature_detected!("avx") |
| 193 |
&& std::arch::is_x86_feature_detected!("fma") |
| 194 |
{ |
| 195 |
secf_fma_impl |
| 196 |
} else { |
| 197 |
secf_gen_impl |
| 198 |
} |
| 199 |
}); |
| 200 |
unsafe { q(x) } |
| 201 |
} |
| 202 |
} |
| 203 |
|
| 204 |
#[cfg(test)] |
| 205 |
mod tests { |
| 206 |
use super::*; |
| 207 |
|
| 208 |
#[test] |
| 209 |
fn test_f_secf() { |
| 210 |
assert_eq!(f_secf(0.0), 1.0); |
| 211 |
assert_eq!(f_secf(0.5), 1.139494); |
| 212 |
assert_eq!(f_secf(-0.5), 1.139494); |
| 213 |
assert_eq!(f_secf(1.5), 14.136833); |
| 214 |
assert_eq!(f_secf(-1.5), 14.136833); |
| 215 |
assert!(f_secf(f32::INFINITY).is_nan()); |
| 216 |
assert!(f_secf(f32::NEG_INFINITY).is_nan()); |
| 217 |
assert!(f_secf(f32::NAN).is_nan()); |
| 218 |
} |
| 219 |
} |