/* * // Copyright (c) Radzivon Bartoshyk 7/2025. 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::{dd_fmla, f_fmla}; use crate::double_double::DoubleDouble; use crate::exponents::{EXP_REDUCE_T0, EXP_REDUCE_T1}; use crate::hyperbolic::acosh::lpoly_xd_generic; #[cold] pub(crate) fn hyperbolic_exp_accurate(x: f64, t: f64, zt: DoubleDouble) -> DoubleDouble { static CH: [(u64, u64); 3] = [ (0x3a16c16bd194535d, 0x3ff0000000000000), (0xba28259d904fd34f, 0x3fe0000000000000), (0x3c653e93e9f26e62, 0x3fc5555555555555), ]; const L2H: f64 = f64::from_bits(0x3f262e42ff000000); const L2L: f64 = f64::from_bits(0x3d0718432a1b0e26); const L2LL: f64 = f64::from_bits(0x3999ff0342542fc3); let dx = x - L2H * t; let mut dxl = L2L * t; let dxll = f_fmla(L2LL, t, dd_fmla(L2L, t, -dxl)); let dxh = dx + dxl; dxl = ((dx - dxh) + dxl) + dxll; let fl0 = f_fmla( dxh, f64::from_bits(0x3f56c16c169400a7), f64::from_bits(0x3f811111113e93e9), ); let fl = dxh * f_fmla(dxh, fl0, f64::from_bits(0x3fa5555555555555)); let mut f = lpoly_xd_generic(DoubleDouble::new(dxl, dxh), CH, fl); f = DoubleDouble::quick_mult(DoubleDouble::new(dxl, dxh), f); f = DoubleDouble::quick_mult(zt, f); let zh = zt.hi + f.hi; let zl = (zt.hi - zh) + f.hi; let uh = zh + zt.lo; let ul = ((zh - uh) + zt.lo) + zl; let vh = uh + f.lo; let vl = ((uh - vh) + f.lo) + ul; DoubleDouble::new(vl, vh) } #[cold] fn as_sinh_zero(x: f64) -> f64 { static CH: [(u64, u64); 5] = [ (0x3c6555555555552f, 0x3fc5555555555555), (0x3c011111115cf00d, 0x3f81111111111111), (0x3b6a0011c925b85c, 0x3f2a01a01a01a01a), (0xbb6b4e2835532bcd, 0x3ec71de3a556c734), (0xbaedefcf17a6ab79, 0x3e5ae64567f54482), ]; let d2x = DoubleDouble::from_exact_mult(x, x); let yw0 = f_fmla( d2x.hi, f64::from_bits(0x3ce95785063cd974), f64::from_bits(0x3d6ae7f36beea815), ); let y2 = d2x.hi * f_fmla(d2x.hi, yw0, f64::from_bits(0x3de6124613aef206)); let mut y1 = lpoly_xd_generic(d2x, CH, y2); y1 = DoubleDouble::quick_mult_f64(y1, x); y1 = DoubleDouble::quick_mult(y1, d2x); // y2 = y1.l let y0 = DoubleDouble::from_exact_add(x, y1.hi); // y0 = y0.hi let mut p = DoubleDouble::from_exact_add(y0.lo, y1.lo); let mut t = p.hi.to_bits(); if (t & 0x000fffffffffffff) == 0 { let w = p.lo.to_bits(); if ((w ^ t) >> 63) != 0 { t = t.wrapping_sub(1); } else { t = t.wrapping_add(1); } p.hi = f64::from_bits(t); } y0.hi + p.hi } /// Hyperbolic sine function /// /// Max ULP 0.5 pub fn f_sinh(x: f64) -> f64 { /* The function sinh(x) is approximated by a minimax polynomial for |x|<0.25. For other arguments the identity sinh(x)=(exp(|x|)-exp(-|x|))/2*copysign(1,x) is used. For |x|<5 both exponents are calculated with slightly higher precision than double. For 5<|x|<36.736801 the exp(-|x|) is small and is calculated with double precision but exp(|x|) is calculated with higher than double precision. For 36.736801<|x|<710.47586 exp(-|x|) becomes too small and only exp(|x|) is calculated. */ const S: f64 = f64::from_bits(0x40b71547652b82fe); let ax = x.abs(); let v0 = dd_fmla(ax, S, f64::from_bits(0x4198000002000000)); let jt = v0.to_bits(); let v = v0.to_bits() & 0xfffffffffc000000; let t = f64::from_bits(v) - f64::from_bits(0x4198000000000000); let ix = ax.to_bits(); let aix = ix; if aix < 0x3fd0000000000000u64 { // |x| < 0x1p-2 if aix < 0x3e57137449123ef7u64 { // |x| < 0x1.7137449123ef7p-26 /* We have underflow exactly when 0 < |x| < 2^-1022: for RNDU, sinh(2^-1022-2^-1074) would round to 2^-1022-2^-1075 with unbounded exponent range */ return dd_fmla(x, f64::from_bits(0x3c80000000000000), x); } const C: [u64; 5] = [ 0x3fc5555555555555, 0x3f81111111111087, 0x3f2a01a01a12e1c3, 0x3ec71de2e415aa36, 0x3e5aed2bff4269e6, ]; let x2 = x * x; let x3 = x2 * x; let x4 = x2 * x2; let pw0 = f_fmla(x2, f64::from_bits(C[3]), f64::from_bits(C[2])); let pw1 = f_fmla(x2, f64::from_bits(C[1]), f64::from_bits(C[0])); let pw2 = f_fmla(x4, f64::from_bits(C[4]), pw0); let p = x3 * f_fmla(x4, pw2, pw1); let e = x3 * f64::from_bits(0x3ca9000000000000); let lb = x + (p - e); let ub = x + (p + e); if lb == ub { return lb; } return as_sinh_zero(x); } if aix > 0x408633ce8fb9f87du64 { // |x| >~ 710.47586 if aix >= 0x7ff0000000000000u64 { return x + x; } // nan Inf return f64::copysign(f64::from_bits(0x7fe0000000000000), x) * 2.0; } let il: i64 = ((jt.wrapping_shl(14)) >> 40) as i64; let jl = -il; let i1 = il & 0x3f; let i0 = (il >> 6) & 0x3f; let ie = il >> 12; let j1 = jl & 0x3f; let j0 = (jl >> 6) & 0x3f; let je = jl >> 12; let mut sp = (1022i64.wrapping_add(ie) as u64).wrapping_shl(52); let sm = (1022i64.wrapping_add(je) as u64).wrapping_shl(52); let sn0 = EXP_REDUCE_T0[i0 as usize]; let sn1 = EXP_REDUCE_T1[i1 as usize]; let t0h = f64::from_bits(sn0.1); let t0l = f64::from_bits(sn0.0); let t1h = f64::from_bits(sn1.1); let t1l = f64::from_bits(sn1.0); let mut th = t0h * t1h; let mut tl = f_fmla(t0h, t1l, t1h * t0l) + dd_fmla(t0h, t1h, -th); const L2H: f64 = f64::from_bits(0x3f262e42ff000000); const L2L: f64 = f64::from_bits(0x3d0718432a1b0e26); let dx = f_fmla(L2L, t, f_fmla(-L2H, t, ax)); let dx2 = dx * dx; let mx = -dx; const CH: [u64; 4] = [ 0x3ff0000000000000, 0x3fe0000000000000, 0x3fc5555555aaaaae, 0x3fa55555551c98c0, ]; let (mut rl, mut rh); let pp0 = f_fmla(dx, f64::from_bits(CH[3]), f64::from_bits(CH[2])); let pp1 = f_fmla(dx, f64::from_bits(CH[1]), f64::from_bits(CH[0])); let pp = dx * f_fmla(dx2, pp0, pp1); if aix > 0x4014000000000000u64 { // |x| > 5 if aix > 0x40425e4f7b2737fau64 { // |x| >~ 36.736801 sp = (1021i64.wrapping_add(ie) as u64).wrapping_shl(52); let mut rh = th; let mut rl = tl + th * pp; rh *= f64::copysign(1., x); rl *= f64::copysign(1., x); let e = 0.11e-18 * th; let lb = rh + (rl - e); let ub = rh + (rl + e); if lb == ub { return (lb * f64::from_bits(sp)) * 2.; } let mut tt = hyperbolic_exp_accurate(ax, t, DoubleDouble::new(tl, th)); tt = DoubleDouble::from_exact_add(tt.hi, tt.lo); th = tt.hi; tl = tt.lo; th *= f64::copysign(1., x); tl *= f64::copysign(1., x); th += tl; th *= 2.; th *= f64::from_bits(sp); return th; } let q0h = f64::from_bits(EXP_REDUCE_T0[j0 as usize].1); let q1h = f64::from_bits(EXP_REDUCE_T1[j1 as usize].1); let mut qh = q0h * q1h; th *= f64::from_bits(sp); tl *= f64::from_bits(sp); qh *= f64::from_bits(sm); let pm0 = f_fmla(mx, f64::from_bits(CH[3]), f64::from_bits(CH[2])); let pm1 = f_fmla(mx, f64::from_bits(CH[1]), f64::from_bits(CH[0])); let pm = mx * f_fmla(dx2, pm0, pm1); let em = f_fmla(qh, pm, qh); rh = th; rl = f_fmla(th, pp, tl - em); rh *= f64::copysign(1., x); rl *= f64::copysign(1., x); let e = 0.09e-18 * rh; let lb = rh + (rl - e); let ub = rh + (rl + e); if lb == ub { return lb; } let tt = hyperbolic_exp_accurate(ax, t, DoubleDouble::new(tl, th)); th = tt.hi; tl = tt.lo; if aix > 0x403f666666666666u64 { rh = th - qh; rl = ((th - rh) - qh) + tl; } else { qh = q0h * q1h; let q0l = f64::from_bits(EXP_REDUCE_T0[j0 as usize].0); let q1l = f64::from_bits(EXP_REDUCE_T1[j1 as usize].0); let mut ql = f_fmla(q0h, q1l, q1h * q0l) + dd_fmla(q0h, q1h, -qh); qh *= f64::from_bits(sm); ql *= f64::from_bits(sm); let qq = hyperbolic_exp_accurate(-ax, -t, DoubleDouble::new(ql, qh)); rh = th - qq.hi; rl = (((th - rh) - qq.hi) - qq.lo) + tl; } } else { let tq0 = EXP_REDUCE_T0[j0 as usize]; let tq1 = EXP_REDUCE_T1[j1 as usize]; let q0h = f64::from_bits(tq0.1); let q0l = f64::from_bits(tq0.0); let q1h = f64::from_bits(tq1.1); let q1l = f64::from_bits(tq1.0); let mut qh = q0h * q1h; let mut ql = f_fmla(q0h, q1l, q1h * q0l) + dd_fmla(q0h, q1h, -qh); th *= f64::from_bits(sp); tl *= f64::from_bits(sp); qh *= f64::from_bits(sm); ql *= f64::from_bits(sm); let pm0 = f_fmla(mx, f64::from_bits(CH[3]), f64::from_bits(CH[2])); let pm1 = f_fmla(mx, f64::from_bits(CH[1]), f64::from_bits(CH[0])); let pm = mx * f_fmla(dx2, pm0, pm1); let fph = th; let fpl = f_fmla(th, pp, tl); let fmh = qh; let fml = f_fmla(qh, pm, ql); rh = fph - fmh; rl = ((fph - rh) - fmh) - fml + fpl; rh *= f64::copysign(1., x); rl *= f64::copysign(1., x); let e = 0.28e-18 * rh; let lb = rh + (rl - e); let ub = rh + (rl + e); if lb == ub { return lb; } let tt = hyperbolic_exp_accurate(ax, t, DoubleDouble::new(tl, th)); let qq = hyperbolic_exp_accurate(-ax, -t, DoubleDouble::new(ql, qh)); rh = tt.hi - qq.hi; rl = ((tt.hi - rh) - qq.hi) - qq.lo + tt.lo; } let r = DoubleDouble::from_exact_add(rh, rl); rh = r.hi; rl = r.lo; rh *= f64::copysign(1., x); rl *= f64::copysign(1., x); rh += rl; rh } #[cfg(test)] mod tests { use super::*; #[test] fn test_f_sinh() { assert_eq!(f_sinh(1.), 1.1752011936438014); } }