/* * // Copyright (c) Radzivon Bartoshyk 8/2025. 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::polyeval::f_estrin_polyeval5; use crate::sin_cosf::sincosf_eval::sincosf_eval; #[inline(always)] fn sincosf_gen_impl(x: f32) -> (f32, f32) { let x_abs = x.to_bits() & 0x7fff_ffffu32; let xd = x as f64; // |x| <= pi/16 if x_abs <= 0x3e49_0fdbu32 { // |x| < 0.000443633 if x_abs < 0x3980_0000u32 { if x_abs == 0u32 { // For signed zeros. return (x, 1.0); } #[cfg(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" ))] { use crate::common::f_fmlaf; let sf = f_fmlaf(x, f32::from_bits(0xb3000000), x); let cf = f_fmlaf(f32::from_bits(x_abs), f32::from_bits(0xb3000000), 1.); return (sf, cf); } #[cfg(not(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" )))] { let sf = f_fmla(xd, f64::from_bits(0xbe60000000000000), xd) as f32; let cf = f_fmla(xd, f64::from_bits(0xbe60000000000000), 1.) as f32; return (sf, cf); } } let xsqr = xd * xd; /* Generated by Sollya: f_sinpi_16 = sin(x)/x; Q = fpminimax(f_sinpi_16, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/16]); See ./notes/sincosf.sollya */ let p = f_estrin_polyeval5( xsqr, f64::from_bits(0x3ff0000000000000), f64::from_bits(0xbfc55555555554c6), f64::from_bits(0x3f81111111085e65), f64::from_bits(0xbf2a019f70fb4d4f), f64::from_bits(0x3ec718d179815e74), ); let sf = (xd * p) as f32; // Cosine // Generated poly by Sollya: // f_cos_16 = cos(x); // // Q = fpminimax(f_cos_16, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/16]); let cf = f_estrin_polyeval5( xsqr, f64::from_bits(0x3ff0000000000000), f64::from_bits(0xbfdffffffffffcea), f64::from_bits(0x3fa55555553d611a), f64::from_bits(0xbf56c16b2e26561a), f64::from_bits(0x3ef9faa67b9da80b), ); return (sf, cf as f32); } if x_abs >= 0x7f80_0000u32 { return (x + f32::NAN, x + f32::NAN); } // Formula: // sin(x) = sin((k + y)*pi/32) // = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..31 are precomputed // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are // computed using degree-7 and degree-6 minimax polynomials generated by // Sollya respectively. let rs = sincosf_eval(xd, x_abs); let sf = f_fmla(rs.sin_y, rs.cos_k, f_fmla(rs.cosm1_y, rs.sin_k, rs.sin_k)) as f32; let cf = f_fmla(rs.sin_y, -rs.sin_k, f_fmla(rs.cosm1_y, rs.cos_k, rs.cos_k)) as f32; (sf, cf) } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] #[target_feature(enable = "avx", enable = "fma")] unsafe fn sincosf_fma_impl(x: f32) -> (f32, f32) { let x_abs = x.to_bits() & 0x7fff_ffffu32; let xd = x as f64; // |x| <= pi/16 if x_abs <= 0x3e49_0fdbu32 { // |x| < 0.000443633 if x_abs < 0x3980_0000u32 { if x_abs == 0u32 { // For signed zeros. return (x, 1.0); } let sf = f32::mul_add(x, f32::from_bits(0xb3000000), x); let cf = f32::mul_add(f32::from_bits(x_abs), f32::from_bits(0xb3000000), 1.); return (sf, cf); } let xsqr = xd * xd; /* Generated by Sollya: f_sinpi_16 = sin(x)/x; Q = fpminimax(f_sinpi_16, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/16]); See ./notes/sincosf.sollya */ use crate::polyeval::d_estrin_polyeval5; let p = d_estrin_polyeval5( xsqr, f64::from_bits(0x3ff0000000000000), f64::from_bits(0xbfc55555555554c6), f64::from_bits(0x3f81111111085e65), f64::from_bits(0xbf2a019f70fb4d4f), f64::from_bits(0x3ec718d179815e74), ); let sf = (xd * p) as f32; // Cosine // Generated poly by Sollya: // f_cos_16 = cos(x); // // Q = fpminimax(f_cos_16, [|0, 2, 4, 6, 8|], [|1, D...|], [0, pi/16]); let cf = d_estrin_polyeval5( xsqr, f64::from_bits(0x3ff0000000000000), f64::from_bits(0xbfdffffffffffcea), f64::from_bits(0x3fa55555553d611a), f64::from_bits(0xbf56c16b2e26561a), f64::from_bits(0x3ef9faa67b9da80b), ); return (sf, cf as f32); } if x_abs >= 0x7f80_0000u32 { return (x + f32::NAN, x + f32::NAN); } // Formula: // sin(x) = sin((k + y)*pi/32) // = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..31 are precomputed // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are // computed using degree-7 and degree-6 minimax polynomials generated by // Sollya respectively. use crate::sin_cosf::sincosf_eval::sincosf_eval_fma; let rs = sincosf_eval_fma(xd, x_abs); let sf = f64::mul_add( rs.sin_y, rs.cos_k, f64::mul_add(rs.cosm1_y, rs.sin_k, rs.sin_k), ) as f32; let cf = f64::mul_add( rs.sin_y, -rs.sin_k, f64::mul_add(rs.cosm1_y, rs.cos_k, rs.cos_k), ) as f32; (sf, cf) } /// Sine and cosine /// /// Max found ULP on working range 0.49999967 #[inline] pub fn f_sincosf(x: f32) -> (f32, f32) { #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] { sincosf_gen_impl(x) } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] { use std::sync::OnceLock; static EXECUTOR: OnceLock (f32, f32)> = OnceLock::new(); let q = EXECUTOR.get_or_init(|| { if std::arch::is_x86_feature_detected!("avx") && std::arch::is_x86_feature_detected!("fma") { sincosf_fma_impl } else { sincosf_gen_impl } }); unsafe { q(x) } } } #[cfg(test)] mod tests { use super::*; #[test] fn f_sincosf_test() { let sincos0 = f_sincosf(0.0); assert!(sincos0.0 < 1e-8); assert_eq!(sincos0.1, 1.0); let sincos_pi = f_sincosf(std::f32::consts::PI); assert!(sincos_pi.0 < 1e-8); let sincos_pi_0_5 = f_sincosf(0.5); assert_eq!(sincos_pi_0_5.0, 0.47942555); assert_eq!(sincos_pi_0_5.1, 0.87758255); let sincos_pi_n0_5 = f_sincosf(-0.5); assert_eq!(sincos_pi_n0_5.0, -0.47942555); assert_eq!(sincos_pi_n0_5.1, 0.87758255); let v_z = f_sincosf(0.0002480338); assert_eq!(v_z.1, 0.9999999692396206); } #[test] fn f_sincosf_edge_test() { let s0 = f_sincosf(f32::INFINITY); assert!(s0.0.is_nan()); assert!(s0.1.is_nan()); let s1 = f_sincosf(f32::NEG_INFINITY); assert!(s1.0.is_nan()); assert!(s1.1.is_nan()); let s2 = f_sincosf(f32::NAN); assert!(s2.0.is_nan()); assert!(s2.1.is_nan()); } }