/* * // Copyright (c) Radzivon Bartoshyk 4/2025. All rights reserved. * // * // Redistribution and use in source and binary forms, with or without modification, * // are permitted provided that the following conditions are met: * // * // 1. Redistributions of source code must retain the above copyright notice, this * // list of conditions and the following disclaimer. * // * // 2. Redistributions in binary form must reproduce the above copyright notice, * // this list of conditions and the following disclaimer in the documentation * // and/or other materials provided with the distribution. * // * // 3. Neither the name of the copyright holder nor the names of its * // contributors may be used to endorse or promote products derived from * // this software without specific prior written permission. * // * // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" * // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE * // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE * // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE * // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL * // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR * // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER * // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, * // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE * // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE. */ use crate::common::{f_fmla, f_fmlaf}; #[inline(always)] fn atanf_gen_impl f64>(x: f32, fma: Q) -> f32 { const PI2: f64 = f64::from_bits(0x3ff921fb54442d18); let t = x.to_bits(); let e = (t >> 23) & 0xff; let gt = e >= 127; let ta = t & 0x7fffffff; if ta >= 0x4c700518u32 { // |x| >= 6.29198e+07 if ta > 0x7f800000u32 { return x + x; } // nan return f32::copysign(PI2 as f32, x); // inf or |x| >= 6.29198e+07 } if e < 127 - 13 { // |x| < 2^-13 if e < 127 - 25 { // |x| < 2^-25 if t << 1 == 0 { return x; } let res = f_fmlaf(-x, x.abs(), x); return res; } return f_fmlaf(-f64::from_bits(0x3fd5555560000000) as f32 * x, x * x, x); } /* now |x| >= 0.00012207 */ let mut z = x as f64; if gt { z = 1.0 / z; } /* gt is non-zero for |x| >= 1 */ let z2 = z * z; let z4 = z2 * z2; let z8 = z4 * z4; /* polynomials generated using rminimax (https://gitlab.inria.fr/sfilip/rminimax) with the following command: ./ratapprox --function="atan(x)" --dom=[0.000122070,1] --num=[x,x^3,x^5,x^7,x^9,x^11,x^13] --den=[1,x^2,x^4,x^6,x^8,x^10,x^12] --output=atanf.sollya --log (see output atanf.sollya) The coefficient cd[0] was slightly reduced from the original value 0.330005 to avoid an exceptional case for |x| = 0.069052 and rounding to nearest. */ const CN: [u64; 7] = [ 0x3fd51eccde075d67, 0x3fea76bb5637f2f2, 0x3fe81e0eed20de88, 0x3fd376c8ca67d11d, 0x3faaec7b69202ac6, 0x3f69561899acc73e, 0x3efbf9fa5b67e600, ]; const CD: [u64; 7] = [ 0x3fd51eccde075d66, 0x3fedfbdd7b392d28, 0x3ff0000000000000, 0x3fdfd22bf0e89b54, 0x3fbd91ff8b576282, 0x3f8653ea99fc9bb0, 0x3f31e7fcc202340a, ]; let mut cn0 = fma(z2, f64::from_bits(CN[1]), f64::from_bits(CN[0])); let cn2 = fma(z2, f64::from_bits(CN[3]), f64::from_bits(CN[2])); let mut cn4 = fma(z2, f64::from_bits(CN[5]), f64::from_bits(CN[4])); let cn6 = f64::from_bits(CN[6]); cn0 = fma(z4, cn2, cn0); cn4 = fma(z4, cn6, cn4); cn0 = fma(z8, cn4, cn0); cn0 *= z; let mut cd0 = fma(z2, f64::from_bits(CD[1]), f64::from_bits(CD[0])); let cd2 = fma(z2, f64::from_bits(CD[3]), f64::from_bits(CD[2])); let mut cd4 = fma(z2, f64::from_bits(CD[5]), f64::from_bits(CD[4])); let cd6 = f64::from_bits(CD[6]); cd0 = fma(z4, cd2, cd0); cd4 = fma(z4, cd6, cd4); cd0 = fma(z8, cd4, cd0); let r = cn0 / cd0; if !gt { return r as f32; } /* for |x| < 1, (float) r is correctly rounded */ const PI_OVER2_H: f64 = f64::from_bits(0x3ff9000000000000); const PI_OVER2_L: f64 = f64::from_bits(0x3f80fdaa22168c23); /* now r approximates atan(1/x), we use atan(x) + atan(1/x) = sign(x)*pi/2, where PI_OVER2_H + PI_OVER2_L approximates pi/2. With sign(z)*L + (-r + sign(z)*H), it fails for x=0x1.98c252p+12 and rounding upward. With sign(z)*PI - r, where PI is a double approximation of pi to nearest, it fails for x=0x1.ddf9f6p+0 and rounding upward. */ ((f64::copysign(PI_OVER2_L, z) - r) + f64::copysign(PI_OVER2_H, z)) as f32 } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] #[target_feature(enable = "avx", enable = "fma")] unsafe fn atanf_fma_impl(x: f32) -> f32 { atanf_gen_impl(x, f64::mul_add) } /// Computes atan /// /// Max found ULP 0.49999973 #[inline] pub fn f_atanf(x: f32) -> f32 { #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] { atanf_gen_impl(x, f_fmla) } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] { use std::sync::OnceLock; static EXECUTOR: OnceLock f32> = OnceLock::new(); let q = EXECUTOR.get_or_init(|| { if std::arch::is_x86_feature_detected!("avx") && std::arch::is_x86_feature_detected!("fma") { atanf_fma_impl } else { fn def_atanf(x: f32) -> f32 { atanf_gen_impl(x, f_fmla) } def_atanf } }); unsafe { q(x) } } } #[cfg(test)] mod tests { use super::*; #[test] fn f_atan_test() { assert!( (f_atanf(1.0) - std::f32::consts::PI / 4f32).abs() < 1e-6, "Invalid result {}", f_atanf(1f32) ); assert!( (f_atanf(2f32) - 1.107148717794090503017065f32).abs() < 1e-6, "Invalid result {}", f_atanf(2f32) ); assert!( (f_atanf(5f32) - 1.3734007669450158608612719264f32).abs() < 1e-6, "Invalid result {}", f_atanf(5f32) ); } }