/* * // Copyright (c) Radzivon Bartoshyk 9/2025. 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; use crate::exponents::core_expdf; use crate::polyeval::{f_estrin_polyeval8, f_polyeval6}; #[inline] fn core_erfcx(x: f32) -> f64 { // x here is already always > 1 let dx = x as f64; if x < 8. { // Rational approximant generated by Wolfram Mathematica: // <75,MaxIterations->100] // num=Numerator[approx]; // den=Denominator[approx]; // coeffs=CoefficientList[num,z]; // TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50},ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] // coeffs=CoefficientList[den,z]; // TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50},ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] let p_num = f_estrin_polyeval8( dx, f64::from_bits(0x3ff00000804c8f8f), f64::from_bits(0x3ffb7307ea8fdbeb), f64::from_bits(0x3ff7081ba7bc735c), f64::from_bits(0x3fe767338b33532a), f64::from_bits(0x3fce3c8288507fd6), f64::from_bits(0x3fa7ca2cb4ae697f), f64::from_bits(0x3f72b11b0dfb2348), f64::from_bits(0xbd9f64f0c15c479b), ); let p_den = f_estrin_polyeval8( dx, f64::from_bits(0x3ff0000000000000), f64::from_bits(0x4006c071e850132e), f64::from_bits(0x400d30326bc347ee), f64::from_bits(0x40060d8d56bada75), f64::from_bits(0x3ff56643fc4580eb), f64::from_bits(0x3fdb0e194e72a513), f64::from_bits(0x3fb5154759b61be3), f64::from_bits(0x3f8090b063cce524), ); return p_num / p_den; } // for large x erfcx(x) ~ 1/sqrt(pi) / x * R(1/x) const ONE_OVER_SQRT_PI: f64 = f64::from_bits(0x3fe20dd750429b6d); let r = 1. / dx; // Rational approximant generated by Wolfram Mathematica: // <75,MaxIterations->100] // num=Numerator[approx][[1]]; // den=Denominator[approx][[1]]; // coeffs=CoefficientList[num,z]; // TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50},ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] // coeffs=CoefficientList[den,z]; // TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50},ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] let p_num = f_polyeval6( r, f64::from_bits(0x3ff0000000000002), f64::from_bits(0xbfd09caf2bb541c3), f64::from_bits(0x40132238367ae454), f64::from_bits(0xc0060bc62c3711b1), f64::from_bits(0x40024a90d229158d), f64::from_bits(0xc0013665d8ff3813), ); let p_den = f_polyeval6( r, f64::from_bits(0x3ff0000000000000), f64::from_bits(0xbfd09caf2bb5101d), f64::from_bits(0x4015223836772f2c), f64::from_bits(0xc00715911b5f5f5c), f64::from_bits(0x4010b66411ec4e1f), f64::from_bits(0xc00b325c767ed436), ); (r * ONE_OVER_SQRT_PI) * (p_num / p_den) } /// Scaled complementary error function (exp(x^2)*erfc(x)) /// /// ulp 0.5 pub fn f_erfcxf(x: f32) -> f32 { let ux = x.to_bits().wrapping_shl(1); if ux >= 0xffu32 << 24 || ux <= 0x6499_999au32 { // |x| == 0, |x| == inf, |x| == NaN, |x| <= 1.19209290e-08 if ux <= 0x6499_999au32 { // |x| == 0, |x| <= 1.19209290e-08 return 1.; } if x.is_infinite() { return if x.is_sign_positive() { 0. } else { f32::INFINITY }; } return f32::NAN; // x == NaN } let ax = x.to_bits() & 0x7fff_ffff; if x <= -9.382415 { // x <= -9.382415 return f32::INFINITY; } if ax <= 0x34000000u32 { // |x| < ulp(1) we use taylor series at 0 // erfcx(x) ~ 1-(2 x)/Sqrt[\[Pi]]+x^2-(4 x^3)/(3 Sqrt[\[Pi]])+x^4/2-(8 x^5)/(15 Sqrt[\[Pi]])+O[x]^6 #[cfg(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" ))] { use crate::common::f_fmlaf; const M_TWO_OVER_SQRT_PI: f32 = f32::from_bits(0xbf906ebb); return f_fmlaf(x, M_TWO_OVER_SQRT_PI, 1.); } #[cfg(not(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" )))] { use crate::common::f_fmla; const M_TWO_OVER_SQRT_PI: f64 = f64::from_bits(0xbff20dd750429b6d); let dx = x as f64; return f_fmla(dx, M_TWO_OVER_SQRT_PI, 1.) as f32; } } if ax <= 0x3f800000u32 { // |x| <= 1 let dx = x as f64; // Generated by Wolfram Mathematica: // <75,MaxIterations->100] // num=Numerator[approx][[1]]; // den=Denominator[approx][[1]]; // coeffs=CoefficientList[num,z]; // TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50},ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] // coeffs=CoefficientList[den,z]; // TableForm[Table[Row[{"'",NumberForm[coeffs[[i+1]],{50,50},ExponentFunction->(Null&)],"',"}],{i,0,Length[coeffs]-1}]] let p_num = f_estrin_polyeval8( dx, f64::from_bits(0x3feffffffffffff8), f64::from_bits(0x3ff26c328bd2dc5f), f64::from_bits(0x3fe6f91b9fa5f58c), f64::from_bits(0x3fd09edf3fcf5ee1), f64::from_bits(0x3faddb3bcedbff91), f64::from_bits(0x3f7e43b5dd4b7587), f64::from_bits(0x3f3baab6b3e61d7b), f64::from_bits(0xbe83e7d629825321), ); let p_den = f_estrin_polyeval8( dx, f64::from_bits(0x3ff0000000000000), f64::from_bits(0x40023d04ee0abc28), f64::from_bits(0x400252b377263d61), f64::from_bits(0x3ff510af7f826479), f64::from_bits(0x3fddfc089c4731ed), f64::from_bits(0x3fba79b040e28b0a), f64::from_bits(0x3f8aea2f3579235a), f64::from_bits(0x3f485d2875b4f88c), ); return (p_num / p_den) as f32; } let erfcx_abs_x = core_erfcx(f32::from_bits(ax)); if x < 0. { // exp(x^2)erfc(-x) = 2*exp(x^2) - erfcx(|x|) let dx = x as f64; return f_fmla(2., core_expdf(dx * dx), -erfcx_abs_x) as f32; } erfcx_abs_x as f32 } #[cfg(test)] mod tests { use super::*; #[test] fn test_erfcx() { assert_eq!(f_erfcxf(5.19209290e-09), 1.0); assert_eq!(f_erfcxf(1.19209290e-08), 1.0); assert_eq!(f_erfcxf(f32::EPSILON), 0.9999999); assert_eq!(f_erfcxf(12.1), 0.046469606); assert_eq!(f_erfcxf(7.1), 0.07869752); assert_eq!(f_erfcxf(1.1), 0.40173045); assert_eq!(f_erfcxf(-0.23), 1.3232007); assert_eq!(f_erfcxf(-1.4325), 15.234794); assert_eq!(f_erfcxf(-10.), f32::INFINITY); assert_eq!(f_erfcxf(f32::INFINITY), 0.); assert_eq!(f_erfcxf(f32::NEG_INFINITY), f32::INFINITY); assert!(f_erfcxf(f32::NAN).is_nan()); } }