/* * // Copyright (c) Radzivon Bartoshyk 6/2025. 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::{dyad_fmla, f_fmla, min_normal_f64}; use crate::double_double::DoubleDouble; use crate::dyadic_float::{DyadicFloat128, DyadicSign}; use crate::rounding::CpuRound; use crate::sin_helper::sincos_eval_dd; use crate::sin_table::SIN_K_PI_OVER_128; use crate::sincos_dyadic::SIN_K_PI_OVER_128_F128; use crate::sincos_reduce::LargeArgumentReduction; // For 2^-7 < |x| < 2^16, return k and u such that: // k = round(x * 128/pi) // x mod pi/128 = x - k * pi/128 ~ u.hi + u.lo // Error bound: // |(x - k * pi/128) - (u_hi + u_lo)| <= max(ulp(ulp(u_hi)), 2^-119) // <= 2^-111. #[inline] pub(crate) fn range_reduction_small(x: f64) -> (DoubleDouble, u64) { const MPI_OVER_128: [u64; 3] = [0xbf9921fb54400000, 0xbd70b4611a600000, 0xbb43198a2e037073]; const ONE_TWENTY_EIGHT_OVER_PI_D: f64 = f64::from_bits(0x40445f306dc9c883); let prod_hi = x * ONE_TWENTY_EIGHT_OVER_PI_D; let kd = prod_hi.cpu_round(); // Let y = x - k * (pi/128) // Then |y| < pi / 256 // With extra rounding errors, we can bound |y| < 1.6 * 2^-7. let y_hi = f_fmla(kd, f64::from_bits(MPI_OVER_128[0]), x); // Exact // |u.hi| < 1.6*2^-7 let u_hi = f_fmla(kd, f64::from_bits(MPI_OVER_128[1]), y_hi); let u0 = y_hi - u_hi; // Exact // |u.lo| <= max(ulp(u.hi), |kd * MPI_OVER_128[2]|) let u1 = f_fmla(kd, f64::from_bits(MPI_OVER_128[1]), u0); // Exact let u_lo = f_fmla(kd, f64::from_bits(MPI_OVER_128[2]), u1); // Error bound: // |x - k * pi/128| - (u.hi + u.lo) <= ulp(u.lo) // <= ulp(max(ulp(u.hi), kd*MPI_OVER_128[2])) // <= 2^(-7 - 104) = 2^-111. (DoubleDouble::new(u_lo, u_hi), unsafe { kd.to_int_unchecked::() as u64 // indeterminate values is always filtered out before this call, as well only lowest bits are used }) } #[inline] pub(crate) fn get_sin_k_rational(kk: u64) -> DyadicFloat128 { let idx = if (kk & 64) != 0 { 64 - (kk & 63) } else { kk & 63 }; let mut ans = SIN_K_PI_OVER_128_F128[idx as usize]; if (kk & 128) != 0 { ans.sign = DyadicSign::Neg; } ans } pub(crate) struct SinCos { pub(crate) v_sin: DoubleDouble, pub(crate) v_cos: DoubleDouble, pub(crate) err: f64, } #[inline] pub(crate) fn sincos_eval(u: DoubleDouble) -> SinCos { // Evaluate sin(y) = sin(x - k * (pi/128)) // We use the degree-7 Taylor approximation: // sin(y) ~ y - y^3/3! + y^5/5! - y^7/7! // Then the error is bounded by: // |sin(y) - (y - y^3/3! + y^5/5! - y^7/7!)| < |y|^9/9! < 2^-54/9! < 2^-72. // For y ~ u_hi + u_lo, fully expanding the polynomial and drop any terms // < ulp(u_hi^3) gives us: // y - y^3/3! + y^5/5! - y^7/7! = ... // ~ u_hi + u_hi^3 * (-1/6 + u_hi^2 * (1/120 - u_hi^2 * 1/5040)) + // + u_lo (1 + u_hi^2 * (-1/2 + u_hi^2 / 24)) let u_hi_sq = u.hi * u.hi; // Error < ulp(u_hi^2) < 2^(-6 - 52) = 2^-58. // p1 ~ 1/120 + u_hi^2 / 5040. let p1 = f_fmla( u_hi_sq, f64::from_bits(0xbf2a01a01a01a01a), f64::from_bits(0x3f81111111111111), ); // q1 ~ -1/2 + u_hi^2 / 24. let q1 = f_fmla( u_hi_sq, f64::from_bits(0x3fa5555555555555), f64::from_bits(0xbfe0000000000000), ); let u_hi_3 = u_hi_sq * u.hi; // p2 ~ -1/6 + u_hi^2 (1/120 - u_hi^2 * 1/5040) let p2 = f_fmla(u_hi_sq, p1, f64::from_bits(0xbfc5555555555555)); // q2 ~ 1 + u_hi^2 (-1/2 + u_hi^2 / 24) let q2 = f_fmla(u_hi_sq, q1, 1.0); let sin_lo = f_fmla(u_hi_3, p2, u.lo * q2); // Overall, |sin(y) - (u_hi + sin_lo)| < 2*ulp(u_hi^3) < 2^-69. // Evaluate cos(y) = cos(x - k * (pi/128)) // We use the degree-8 Taylor approximation: // cos(y) ~ 1 - y^2/2 + y^4/4! - y^6/6! + y^8/8! // Then the error is bounded by: // |cos(y) - (...)| < |y|^10/10! < 2^-81 // For y ~ u_hi + u_lo, fully expanding the polynomial and drop any terms // < ulp(u_hi^3) gives us: // 1 - y^2/2 + y^4/4! - y^6/6! + y^8/8! = ... // ~ 1 - u_hi^2/2 + u_hi^4(1/24 + u_hi^2 (-1/720 + u_hi^2/40320)) + // + u_hi u_lo (-1 + u_hi^2/6) // We compute 1 - u_hi^2 accurately: // v_hi + v_lo ~ 1 - u_hi^2/2 // with error <= 2^-105. let u_hi_neg_half = (-0.5) * u.hi; let (mut v_lo, v_hi); #[cfg(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" ))] { v_hi = f_fmla(u.hi, u_hi_neg_half, 1.0); v_lo = 1.0 - v_hi; // Exact v_lo = f_fmla(u.hi, u_hi_neg_half, v_lo); } #[cfg(not(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" )))] { let u_hi_sq_neg_half = DoubleDouble::from_exact_mult(u.hi, u_hi_neg_half); let v = DoubleDouble::from_exact_add(1.0, u_hi_sq_neg_half.hi); v_lo = v.lo; v_lo += u_hi_sq_neg_half.lo; v_hi = v.hi; } // r1 ~ -1/720 + u_hi^2 / 40320 let r1 = f_fmla( u_hi_sq, f64::from_bits(0x3efa01a01a01a01a), f64::from_bits(0xbf56c16c16c16c17), ); // s1 ~ -1 + u_hi^2 / 6 let s1 = f_fmla(u_hi_sq, f64::from_bits(0x3fc5555555555555), -1.0); let u_hi_4 = u_hi_sq * u_hi_sq; let u_hi_u_lo = u.hi * u.lo; // r2 ~ 1/24 + u_hi^2 (-1/720 + u_hi^2 / 40320) let r2 = f_fmla(u_hi_sq, r1, f64::from_bits(0x3fa5555555555555)); // s2 ~ v_lo + u_hi * u_lo * (-1 + u_hi^2 / 6) let s2 = f_fmla(u_hi_u_lo, s1, v_lo); let cos_lo = f_fmla(u_hi_4, r2, s2); // Overall, |cos(y) - (v_hi + cos_lo)| < 2*ulp(u_hi^4) < 2^-75. let sin_u = DoubleDouble::from_exact_add(u.hi, sin_lo); let cos_u = DoubleDouble::from_exact_add(v_hi, cos_lo); let err = f_fmla( u_hi_3, f64::from_bits(0x3cc0000000000000), f64::from_bits(0x3960000000000000), ); SinCos { v_sin: sin_u, v_cos: cos_u, err, } } #[cold] #[inline(never)] fn sin_accurate(y: DoubleDouble, sin_k: DoubleDouble, cos_k: DoubleDouble) -> f64 { let r_sincos = sincos_eval_dd(y); // k is an integer and -pi / 256 <= y <= pi / 256. // Then sin(x) = sin((k * pi/128 + y) // = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128) let sin_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, sin_k); let cos_k_sin_y = DoubleDouble::quick_mult(r_sincos.v_sin, cos_k); let rr = DoubleDouble::full_dd_add(sin_k_cos_y, cos_k_sin_y); rr.to_f64() } #[cold] pub(crate) fn sin_accurate_near_zero(x: f64) -> f64 { // Generated by Sollya: // d = [0, 0.0490874]; // f_sin = sin(y)/y; // Q = fpminimax(f_sin, [|0, 2, 4, 6, 8, 10, 12|], [|107...|], d, relative, floating); const C: [(u64, u64); 7] = [ (0x0000000000000000, 0x3ff0000000000000), (0xbc65555555555217, 0xbfc5555555555555), (0x3c011110f7d49e8c, 0x3f81111111111111), (0xbb65e5b30986fc80, 0xbf2a01a01a01a01a), (0xbb689e86245bd566, 0x3ec71de3a556c6e5), (0xbaccb3d55ccfca58, 0xbe5ae64567954935), (0x3a6333ef7a00ce40, 0x3de6120ca00ce3a2), ]; let x2 = DoubleDouble::from_exact_mult(x, x); let mut p = DoubleDouble::quick_mul_add( x2, DoubleDouble::from_bit_pair(C[6]), DoubleDouble::from_bit_pair(C[5]), ); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[4])); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); p = DoubleDouble::quick_mul_add_f64(x2, p, f64::from_bits(0x3ff0000000000000)); p = DoubleDouble::quick_mult_f64(p, x); p.to_f64() } /// Sine for double precision /// /// ULP 0.5 pub fn f_sin(x: f64) -> f64 { let x_e = (x.to_bits() >> 52) & 0x7ff; const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; let y: DoubleDouble; let k; let mut argument_reduction = LargeArgumentReduction::default(); if x_e < E_BIAS + 16 { // |x| < 2^32 (with FMA) or |x| < 2^23 (w/o FMA) let ax = x.to_bits() & 0x7fff_ffff_ffff_ffff; if ax <= 0x3fa921fbd34a9550 { // |x| <= 0.0490874 if x_e < E_BIAS - 26 { // |x| < 2^-26 if x == 0.0 { // Signed zeros. return x; } // For |x| < 2^-26, |sin(x) - x| < ulp(x)/2. return dyad_fmla(x, f64::from_bits(0xbc90000000000000), x); } // Polynomial for sin(x)/x // Generated by Sollya: // d = [0, 0.0490874]; // f_sin = sin(y)/y; // Q = fpminimax(f_sin, [|0, 2, 4, 6, 8|], [|1, D...|], d, relative, floating); let x2 = x * x; let x4 = x2 * x2; const C: [u64; 4] = [ 0xbfc5555555555555, 0x3f8111111110e45a, 0xbf2a019ffd7fdaaf, 0x3ec71819b9bf01ef, ]; let p01 = f_fmla(x2, f64::from_bits(C[1]), f64::from_bits(C[0])); let p23 = f_fmla(x2, f64::from_bits(C[3]), f64::from_bits(C[2])); let w0 = f_fmla(x4, p23, p01); let w1 = x2 * w0 * x; let r = DoubleDouble::from_exact_add(x, w1); let err = f_fmla( x2, f64::from_bits(0x3cb0000000000000), // 2^-52 f64::from_bits(0x3be0000000000000), // 2^-65 ); let ub = r.hi + (r.lo + err); let lb = r.hi + (r.lo - err); if ub == lb { return ub; } return sin_accurate_near_zero(x); } // Small range reduction. (y, k) = range_reduction_small(x); } else { // Inf or NaN if x_e > 2 * E_BIAS { // sin(+-Inf) = NaN return x + f64::NAN; } // Large range reduction. (k, y) = argument_reduction.reduce(x); } let r_sincos = sincos_eval(y); // cos(k * pi/128) = sin(k * pi/128 + pi/2) = sin((k + 64) * pi/128). let sk = SIN_K_PI_OVER_128[(k & 255) as usize]; let ck = SIN_K_PI_OVER_128[((k.wrapping_add(64)) & 255) as usize]; let sin_k = DoubleDouble::from_bit_pair(sk); let cos_k = DoubleDouble::from_bit_pair(ck); let sin_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, sin_k); let cos_k_sin_y = DoubleDouble::quick_mult(r_sincos.v_sin, cos_k); // sin_k_cos_y is always >> cos_k_sin_y let mut rr = DoubleDouble::from_exact_add(sin_k_cos_y.hi, cos_k_sin_y.hi); rr.lo += sin_k_cos_y.lo + cos_k_sin_y.lo; let rlp = rr.lo + r_sincos.err; let rlm = rr.lo - r_sincos.err; let r_upper = rr.hi + rlp; // (rr.lo + ERR); let r_lower = rr.hi + rlm; // (rr.lo - ERR); // Ziv's accuracy test if r_upper == r_lower { return rr.to_f64(); } sin_accurate(y, sin_k, cos_k) } #[cold] #[inline(never)] fn cos_accurate(y: DoubleDouble, msin_k: DoubleDouble, cos_k: DoubleDouble) -> f64 { let r_sincos = sincos_eval_dd(y); // After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128). // So k is an integer and -pi / 256 <= y <= pi / 256. // Then sin(x) = sin((k * pi/128 + y) // = sin(y) * cos(k*pi/128) + cos(y) * sin(k*pi/128) let sin_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, cos_k); let cos_k_sin_y = DoubleDouble::quick_mult(r_sincos.v_sin, msin_k); let rr = DoubleDouble::full_dd_add(sin_k_cos_y, cos_k_sin_y); rr.to_f64() } #[cold] pub(crate) fn cos_accurate_near_zero(x: f64) -> f64 { // Generated by Sollya: // d = [0, 0.0490874]; // f_sin = sin(y)/y; // Q = fpminimax(f_sin, [|0, 2, 4, 6, 8, 10|], [|107...|], d, relative, floating); const C: [(u64, u64); 6] = [ (0xba2fa1c000000000, 0x3ff0000000000000), (0x3b1cd6ead3800000, 0xbfe0000000000000), (0x3c45112931063916, 0x3fa5555555555555), (0xbba1c1e3ab3b09e0, 0xbf56c16c16c16b2b), (0x3b937bc2f4ea7db6, 0x3efa01a019495fca), (0x3b307fd5e1b021b3, 0xbe927e0d57d894f8), ]; let x2 = DoubleDouble::from_exact_mult(x, x); let mut p = DoubleDouble::quick_mul_add( x2, DoubleDouble::from_bit_pair(C[5]), DoubleDouble::from_bit_pair(C[4]), ); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); p = DoubleDouble::quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[0])); p.to_f64() } /// Cosine for double precision /// /// ULP 0.5 pub fn f_cos(x: f64) -> f64 { let x_e = (x.to_bits() >> 52) & 0x7ff; const E_BIAS: u64 = (1u64 << (11 - 1u64)) - 1u64; let y: DoubleDouble; let k; let mut argument_reduction = LargeArgumentReduction::default(); if x_e < E_BIAS + 16 { // |x| < 2^32 (with FMA) or |x| < 2^23 (w/o FMA) let ax = x.to_bits() & 0x7fff_ffff_ffff_ffff; if ax <= 0x3fa921fbd34a9550 { // |x| <= 0.0490874 if x_e < E_BIAS - 27 { // |x| < 2^-27 if x == 0.0 { // Signed zeros. return 1.0; } // For |x| < 2^-26, |sin(x) - x| < ulp(x)/2. return 1.0 - min_normal_f64(); } // Polynomial for cos(x) // Generated by Sollya: // d = [0, 0.0490874]; // f_cos = cos(y); // Q = fpminimax(f_cos, [|0, 2, 4, 6, 8|], [|1, D...|], d, relative, floating); let x2 = x * x; let x4 = x2 * x2; const C: [u64; 4] = [ 0xbfe0000000000000, 0x3fa55555555554a4, 0xbf56c16c1619b84a, 0x3efa013d3d01cf7f, ]; let p01 = f_fmla(x2, f64::from_bits(C[1]), f64::from_bits(C[0])); let p23 = f_fmla(x2, f64::from_bits(C[3]), f64::from_bits(C[2])); let w0 = f_fmla(x4, p23, p01); let w1 = x2 * w0; let r = DoubleDouble::from_exact_add(1., w1); let err = f_fmla( x2, f64::from_bits(0x3cb0000000000000), // 2^-52 f64::from_bits(0x3be0000000000000), // 2^-65 ); let ub = r.hi + (r.lo + err); let lb = r.hi + (r.lo - err); if ub == lb { return ub; } return cos_accurate_near_zero(x); } else { // // Small range reduction. (y, k) = range_reduction_small(x); } } else { // Inf or NaN if x_e > 2 * E_BIAS { // cos(+-Inf) = NaN return x + f64::NAN; } // Large range reduction. (k, y) = argument_reduction.reduce(x); } let r_sincos = sincos_eval(y); // After range reduction, k = round(x * 128 / pi) and y = x - k * (pi / 128). // So k is an integer and -pi / 256 <= y <= pi / 256. // Then cos(x) = cos((k * pi/128 + y) // = cos(y) * cos(k*pi/128) - sin(y) * sin(k*pi/128) let sk = SIN_K_PI_OVER_128[(k.wrapping_add(128) & 255) as usize]; let ck = SIN_K_PI_OVER_128[((k.wrapping_add(64)) & 255) as usize]; let msin_k = DoubleDouble::from_bit_pair(sk); let cos_k = DoubleDouble::from_bit_pair(ck); let cos_k_cos_y = DoubleDouble::quick_mult(r_sincos.v_cos, cos_k); let cos_k_msin_y = DoubleDouble::quick_mult(r_sincos.v_sin, msin_k); // cos_k_cos_y is always >> cos_k_msin_y let mut rr = DoubleDouble::from_exact_add(cos_k_cos_y.hi, cos_k_msin_y.hi); rr.lo += cos_k_cos_y.lo + cos_k_msin_y.lo; let rlp = rr.lo + r_sincos.err; let rlm = rr.lo - r_sincos.err; let r_upper = rr.hi + rlp; // (rr.lo + ERR); let r_lower = rr.hi + rlm; // (rr.lo - ERR); // Ziv's accuracy test if r_upper == r_lower { return rr.to_f64(); } cos_accurate(y, msin_k, cos_k) } #[cfg(test)] mod tests { use super::*; #[test] fn cos_test() { assert_eq!(f_cos(0.0), 1.0); assert_eq!(f_cos(0.0024312312), 0.9999970445588818); assert_eq!(f_cos(1.0), 0.5403023058681398); assert_eq!(f_cos(-0.5), 0.8775825618903728); assert!(f_cos(f64::INFINITY).is_nan()); assert!(f_cos(f64::NEG_INFINITY).is_nan()); assert!(f_cos(f64::NAN).is_nan()); } #[test] fn sin_test() { assert_eq!(f_sin(0.00000002149119332890786), 0.00000002149119332890786); assert_eq!(f_sin(2.8477476437362989E-306), 2.8477476437362989E-306); assert_eq!(f_sin(0.0024312312), 0.002431228804879309); assert_eq!(f_sin(0.0), 0.0); assert_eq!(f_sin(1.0), 0.8414709848078965); assert_eq!(f_sin(-0.5), -0.479425538604203); assert!(f_sin(f64::INFINITY).is_nan()); assert!(f_sin(f64::NEG_INFINITY).is_nan()); assert!(f_sin(f64::NAN).is_nan()); } }