/* * // 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::double_double::DoubleDouble; use crate::polyeval::f_estrin_polyeval5; use crate::sincospi::{GenSinCosPiBackend, SinCosPiBackend, reduce_pi_64}; use crate::sincospi_tables::SINPI_K_PI_OVER_64; /** Sinpi on range [0.0, 0.03515625] Generated poly by Sollya: ```text d = [0, 0.03515625]; f_sincpi = sin(y*pi)/(y*pi); Q = fpminimax(f_sincpi, [|0, 2, 4, 6, 8, 10, 12|], [|107...|], d, relative, floating); ``` See ./notes/sincpi_at_zero_dd.sollya **/ #[cold] fn as_sincpi_zero(x: f64, backend: &B) -> f64 { const C: [(u64, u64); 7] = [ (0xb9d3080000000000, 0x3ff0000000000000), (0xbc81873d86314302, 0xbffa51a6625307d3), (0x3c84871b4ffeefae, 0x3fe9f9cb402bc46c), (0xbc5562d6ae037010, 0xbfc86a8e4720db66), (0xbc386c93f4549bac, 0x3f9ac6805cf31ffd), (0x3c0dbda368edfa40, 0xbf633816a3399d4e), (0xbbcf22ccc18f27a9, 0x3f23736e6a59edd9), ]; let x2 = backend.exact_mult(x, x); let mut p = backend.quick_mul_add( x2, DoubleDouble::from_bit_pair(C[6]), DoubleDouble::from_bit_pair(C[5]), ); p = backend.quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[4])); p = backend.quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[3])); p = backend.quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[2])); p = backend.quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[1])); p = backend.quick_mul_add(x2, p, DoubleDouble::from_bit_pair(C[0])); p.to_f64() } #[inline(always)] fn sincpi_gen_impl(x: f64) -> f64 { let ix = x.to_bits(); let ax = ix & 0x7fff_ffff_ffff_ffff; if ax == 0 { return 1.; } let e: i32 = (ax >> 52) as i32; let m0 = (ix & 0x000fffffffffffff) | (1u64 << 52); let sgn: i64 = (ix as i64) >> 63; let m = ((m0 as i64) ^ sgn).wrapping_sub(sgn); let mut s: i32 = 1063i32.wrapping_sub(e); if s < 0 { if e == 0x7ff { if (ix << 12) == 0 { return f64::NAN; } return x + x; // case x=NaN } s = -s - 1; if s > 10 { return f64::copysign(0.0, x); } let iq: u64 = (m as u64).wrapping_shl(s as u32); if (iq & 2047) == 0 { return f64::copysign(0.0, x); } } if ax <= 0x3fa2000000000000u64 { // |x| <= 0.03515625 if ax < 0x3c90000000000000u64 { // |x| < f64::EPSILON if ax <= 0x3b05798ee2308c3au64 { // |x| <= 2.2204460492503131e-24 return 1.; } // Small values approximated with Taylor poly // sincpi(x) ~ 1 - x^2*Pi^2/6 + O(x^4) #[cfg(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" ))] { const M_SQR_PI_OVER_6: f64 = f64::from_bits(0xbffa51a6625307d3); let p = f_fmla(x, M_SQR_PI_OVER_6 * x, 1.); return p; } #[cfg(not(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" )))] { use crate::common::min_normal_f64; return 1. - min_normal_f64(); } } // Poly generated by Sollya: // d = [0, 0.03515625]; // f_sincpi = sin(y*pi)/(y*pi); // Q = fpminimax(f_sincpi, [|0, 2, 4, 6, 8, 10|], [|107, D...|], d, relative, floating); // See ./notes/sincpi_at_zero.sollya let x2 = x * x; let eps = x * f_fmla( x2, f64::from_bits(0x3d00000000000000), // 2^-47 f64::from_bits(0x3bd0000000000000), // 2^-66 ); const C: [u64; 5] = [ 0xbffa51a6625307d3, 0x3fe9f9cb402bbeaa, 0xbfc86a8e466bbb5b, 0x3f9ac66d887e2f38, 0xbf628473a38d289a, ]; const F: DoubleDouble = DoubleDouble::from_bit_pair((0xbb93f0a925810000, 0x3ff0000000000000)); let p = f_estrin_polyeval5( x2, f64::from_bits(C[0]), f64::from_bits(C[1]), f64::from_bits(C[2]), f64::from_bits(C[3]), f64::from_bits(C[4]), ); let v = DoubleDouble::from_exact_mult(p, x2); let z = DoubleDouble::add(F, v); let lb = z.hi + (z.lo - eps); let ub = z.hi + (z.lo + eps); if lb == ub { return lb; } return as_sincpi_zero(x, &GenSinCosPiBackend {}); } let si = e.wrapping_sub(1011); if si >= 0 && (m0.wrapping_shl(si.wrapping_add(1) as u32)) == 0 { // x is integer or half-integer if (m0.wrapping_shl(si as u32)) == 0 { return f64::copysign(0.0, x); // x is integer } let t = (m0.wrapping_shl((si - 1) as u32)) >> 63; // t = 0 if |x| = 1/2 mod 2, t = 1 if |x| = 3/2 mod 2 #[cfg(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" ))] { let num = if t == 0 { f64::copysign(1.0, x) } else { -f64::copysign(1.0, x) }; const PI: DoubleDouble = DoubleDouble::from_bit_pair((0x3ca1a62633145c07, 0x400921fb54442d18)); let r = DoubleDouble::quick_mult_f64(PI, x); let v = DoubleDouble::from_f64_div_dd(num, r); return v.to_f64(); } #[cfg(not(any( all( any(target_arch = "x86", target_arch = "x86_64"), target_feature = "fma" ), target_arch = "aarch64" )))] { use crate::double_double::two_product_compatible; return if two_product_compatible(x) { let num = if t == 0 { f64::copysign(1.0, x) } else { -f64::copysign(1.0, x) }; const PI: DoubleDouble = DoubleDouble::from_bit_pair((0x3ca1a62633145c07, 0x400921fb54442d18)); let r = DoubleDouble::quick_mult_f64(PI, x); let v = DoubleDouble::from_f64_div_dd(num, r); v.to_f64() } else { use crate::dyadic_float::{DyadicFloat128, DyadicSign}; let num = DyadicFloat128::new_from_f64(if t == 0 { f64::copysign(1.0, x) } else { -f64::copysign(1.0, x) }); const PI: DyadicFloat128 = DyadicFloat128 { sign: DyadicSign::Pos, exponent: -126, mantissa: 0xc90fdaa2_2168c234_c4c6628b_80dc1cd1_u128, }; let dx = DyadicFloat128::new_from_f64(x); let r = (PI * dx).reciprocal(); (num * r).fast_as_f64() }; } } let (y, k) = reduce_pi_64(x); // cos(k * pi/64) = sin(k * pi/64 + pi/2) = sin((k + 32) * pi/64). let sin_k = DoubleDouble::from_bit_pair(SINPI_K_PI_OVER_64[((k as u64) & 127) as usize]); let cos_k = DoubleDouble::from_bit_pair( SINPI_K_PI_OVER_64[((k as u64).wrapping_add(32) & 127) as usize], ); let r_sincos = crate::sincospi::sincospi_eval(y, &GenSinCosPiBackend {}); const PI: DoubleDouble = DoubleDouble::from_bit_pair((0x3ca1a62633145c07, 0x400921fb54442d18)); let scale = DoubleDouble::quick_mult_f64(PI, x); let sin_k_cos_y = DoubleDouble::quick_mult(sin_k, r_sincos.v_cos); let cos_k_sin_y = DoubleDouble::quick_mult(cos_k, r_sincos.v_sin); // 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; rr = DoubleDouble::div(rr, scale); let ub = rr.hi + (rr.lo + r_sincos.err); // (rr.lo + ERR); let lb = rr.hi + (rr.lo - r_sincos.err); // (rr.lo - ERR); if ub == lb { return rr.to_f64(); } sincpi_dd(y, sin_k, cos_k, scale, &GenSinCosPiBackend {}) } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] #[target_feature(enable = "avx", enable = "fma")] unsafe fn sincpi_fma_impl(x: f64) -> f64 { use crate::sincospi::FmaSinCosPiBackend; let ix = x.to_bits(); let ax = ix & 0x7fff_ffff_ffff_ffff; if ax == 0 { return 1.; } let e: i32 = (ax >> 52) as i32; let m0 = (ix & 0x000fffffffffffff) | (1u64 << 52); let sgn: i64 = (ix as i64) >> 63; let m = ((m0 as i64) ^ sgn).wrapping_sub(sgn); let mut s: i32 = 1063i32.wrapping_sub(e); if s < 0 { if e == 0x7ff { if (ix << 12) == 0 { return f64::NAN; } return x + x; // case x=NaN } s = -s - 1; if s > 10 { return f64::copysign(0.0, x); } let iq: u64 = (m as u64).wrapping_shl(s as u32); if (iq & 2047) == 0 { return f64::copysign(0.0, x); } } if ax <= 0x3fa2000000000000u64 { // |x| <= 0.03515625 if ax < 0x3c90000000000000u64 { // |x| < f64::EPSILON if ax <= 0x3b05798ee2308c3au64 { // |x| <= 2.2204460492503131e-24 return 1.; } // Small values approximated with Taylor poly // sincpi(x) ~ 1 - x^2*Pi^2/6 + O(x^4) const M_SQR_PI_OVER_6: f64 = f64::from_bits(0xbffa51a6625307d3); return f64::mul_add(x, M_SQR_PI_OVER_6 * x, 1.); } // Poly generated by Sollya: // d = [0, 0.03515625]; // f_sincpi = sin(y*pi)/(y*pi); // Q = fpminimax(f_sincpi, [|0, 2, 4, 6, 8, 10|], [|107, D...|], d, relative, floating); // See ./notes/sincpi_at_zero.sollya let x2 = x * x; let eps = x * f64::mul_add( x2, f64::from_bits(0x3d00000000000000), // 2^-47 f64::from_bits(0x3bd0000000000000), // 2^-66 ); const C: [u64; 5] = [ 0xbffa51a6625307d3, 0x3fe9f9cb402bbeaa, 0xbfc86a8e466bbb5b, 0x3f9ac66d887e2f38, 0xbf628473a38d289a, ]; const F: DoubleDouble = DoubleDouble::from_bit_pair((0xbb93f0a925810000, 0x3ff0000000000000)); use crate::polyeval::d_estrin_polyeval5; let p = d_estrin_polyeval5( x2, f64::from_bits(C[0]), f64::from_bits(C[1]), f64::from_bits(C[2]), f64::from_bits(C[3]), f64::from_bits(C[4]), ); let v = DoubleDouble::from_exact_mult_fma(p, x2); let z = DoubleDouble::add(F, v); let lb = z.hi + (z.lo - eps); let ub = z.hi + (z.lo + eps); if lb == ub { return lb; } return as_sincpi_zero(x, &FmaSinCosPiBackend {}); } let si = e.wrapping_sub(1011); if si >= 0 && (m0.wrapping_shl(si.wrapping_add(1) as u32)) == 0 { // x is integer or half-integer if (m0.wrapping_shl(si as u32)) == 0 { return f64::copysign(0.0, x); // x is integer } let t = (m0.wrapping_shl((si - 1) as u32)) >> 63; // t = 0 if |x| = 1/2 mod 2, t = 1 if |x| = 3/2 mod 2 let num = if t == 0 { f64::copysign(1.0, x) } else { -f64::copysign(1.0, x) }; const PI: DoubleDouble = DoubleDouble::from_bit_pair((0x3ca1a62633145c07, 0x400921fb54442d18)); let r = DoubleDouble::quick_mult_f64_fma(PI, x); let v = DoubleDouble::from_f64_div_dd_fma(num, r); return v.to_f64(); } use crate::sincospi::reduce_pi_64_fma; let (y, k) = reduce_pi_64_fma(x); // cos(k * pi/64) = sin(k * pi/64 + pi/2) = sin((k + 32) * pi/64). let sin_k = DoubleDouble::from_bit_pair(SINPI_K_PI_OVER_64[((k as u64) & 127) as usize]); let cos_k = DoubleDouble::from_bit_pair( SINPI_K_PI_OVER_64[((k as u64).wrapping_add(32) & 127) as usize], ); let r_sincos = crate::sincospi::sincospi_eval(y, &FmaSinCosPiBackend {}); const PI: DoubleDouble = DoubleDouble::from_bit_pair((0x3ca1a62633145c07, 0x400921fb54442d18)); let scale = DoubleDouble::quick_mult_f64_fma(PI, x); let sin_k_cos_y = DoubleDouble::quick_mult_fma(sin_k, r_sincos.v_cos); let cos_k_sin_y = DoubleDouble::quick_mult_fma(cos_k, r_sincos.v_sin); // 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; rr = DoubleDouble::div_fma(rr, scale); let ub = rr.hi + (rr.lo + r_sincos.err); // (rr.lo + ERR); let lb = rr.hi + (rr.lo - r_sincos.err); // (rr.lo - ERR); if ub == lb { return rr.to_f64(); } sincpi_dd(y, sin_k, cos_k, scale, &FmaSinCosPiBackend {}) } /// Computes sin(PI\*x)/(PI\*x) /// /// Produces normalized sinc. /// /// Max ULP 0.5 pub fn f_sincpi(x: f64) -> f64 { #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] { sincpi_gen_impl(x) } #[cfg(any(target_arch = "x86", target_arch = "x86_64"))] { use std::sync::OnceLock; static EXECUTOR: OnceLock f64> = OnceLock::new(); let q = EXECUTOR.get_or_init(|| { if std::arch::is_x86_feature_detected!("avx") && std::arch::is_x86_feature_detected!("fma") { sincpi_fma_impl } else { fn def_sincpi(x: f64) -> f64 { sincpi_gen_impl(x) } def_sincpi } }); unsafe { q(x) } } } #[cold] #[inline(always)] fn sincpi_dd( x: f64, sin_k: DoubleDouble, cos_k: DoubleDouble, scale: DoubleDouble, backend: &B, ) -> f64 { let r_sincos = crate::sincospi::sincospi_eval_dd(x, backend); let cos_k_sin_y = backend.quick_mult(cos_k, r_sincos.v_sin); let mut rr = backend.mul_add(sin_k, r_sincos.v_cos, cos_k_sin_y); rr = backend.div(rr, scale); rr.to_f64() } #[cfg(test)] mod tests { use super::*; #[test] fn test_sincpi_zero() { assert_eq!(f_sincpi(2.2204460492503131e-24), 1.0); assert_eq!(f_sincpi(f64::EPSILON), 1.0); assert_eq!(f_sincpi(0.007080019335262543), 0.9999175469662566); assert_eq!(f_sincpi(0.05468860710998057), 0.9950875152844803); assert_eq!(f_sincpi(0.5231231231), 0.6068750737806441); assert_eq!(f_sincpi(1.), 0.); assert_eq!(f_sincpi(-1.), 0.); assert_eq!(f_sincpi(-2.), 0.); assert_eq!(f_sincpi(-3.), 0.); assert!(f_sincpi(f64::INFINITY).is_nan()); assert!(f_sincpi(f64::NEG_INFINITY).is_nan()); assert!(f_sincpi(f64::NAN).is_nan()); } }