pxfm/sin_cosf/
cscf.rs

1/*
2 * // Copyright (c) Radzivon Bartoshyk 8/2025. All rights reserved.
3 * //
4 * // Redistribution and use in source and binary forms, with or without modification,
5 * // are permitted provided that the following conditions are met:
6 * //
7 * // 1.  Redistributions of source code must retain the above copyright notice, this
8 * // list of conditions and the following disclaimer.
9 * //
10 * // 2.  Redistributions in binary form must reproduce the above copyright notice,
11 * // this list of conditions and the following disclaimer in the documentation
12 * // and/or other materials provided with the distribution.
13 * //
14 * // 3.  Neither the name of the copyright holder nor the names of its
15 * // contributors may be used to endorse or promote products derived from
16 * // this software without specific prior written permission.
17 * //
18 * // THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
19 * // AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
20 * // IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
21 * // DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
22 * // FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
23 * // DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
24 * // SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
25 * // CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
26 * // OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
27 * // OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
28 */
29use crate::common::f_fmla;
30use crate::polyeval::{f_polyeval3, f_polyeval5};
31use crate::sin_cosf::sincosf_eval::sincosf_eval;
32
33#[inline(always)]
34fn cscf_gen_impl(x: f32) -> f32 {
35    let x_abs = x.to_bits() & 0x7fff_ffffu32;
36    let xd = x as f64;
37
38    // |x| <= pi/16
39    if x_abs <= 0x3e49_0fdbu32 {
40        // |x| < 0.000443633
41        if x_abs < 0x39e8_9769u32 {
42            if x_abs == 0u32 {
43                // For signed zeros.
44                return if x.is_sign_negative() {
45                    f32::NEG_INFINITY
46                } else {
47                    f32::INFINITY
48                };
49            }
50            let dx = x as f64;
51            let c_term = 1. / dx;
52            let x2 = dx * dx;
53            // Maclaurin series
54            // 1/x + x/6 + (7 x^3)/360 + (31 x^5)/15120 + O(x^7)
55            let p = f_polyeval3(
56                x2,
57                f64::from_bits(0x3fc5555555555555),
58                f64::from_bits(0x3f93e93e93e93e94),
59                f64::from_bits(0x3f60b2463814bc5f),
60            );
61            return f_fmla(dx, p, c_term) as f32;
62        }
63
64        let xsqr = xd * xd;
65
66        /*
67        Generated by Sollya:
68        f = 1 / sin(x) - 1/x;
69
70        d = [0.000443633; pi/16];
71        pf = fpminimax(f, [|1, 3, 5, 7, 9|], [|D...|], d, relative, floating);
72
73        See ./notes/cscf.sollya
74         */
75
76        let p = f_polyeval5(
77            xsqr,
78            f64::from_bits(0x3fc5555555555562),
79            f64::from_bits(0x3f93e93e93e730a3),
80            f64::from_bits(0x3f60cbb77382ae6f),
81            f64::from_bits(0x3f2b85bfd4188934),
82            f64::from_bits(0x3ef697a32ebe822d),
83        );
84        return f_fmla(xd, p, 1. / xd) as f32;
85    }
86
87    if x_abs >= 0x7f80_0000u32 {
88        return x + f32::NAN;
89    }
90
91    // Formula:
92    //   sin(x) = sin((k + y)*pi/32)
93    //          = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32)
94    // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..31 are precomputed
95    // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
96    // computed using degree-7 and degree-6 minimax polynomials generated by
97    // Sollya respectively.
98
99    let rs = sincosf_eval(xd, x_abs);
100    (1. / f_fmla(rs.sin_y, rs.cos_k, f_fmla(rs.cosm1_y, rs.sin_k, rs.sin_k))) as f32
101}
102
103#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
104#[target_feature(enable = "avx", enable = "fma")]
105unsafe fn cscf_fma_impl(x: f32) -> f32 {
106    let x_abs = x.to_bits() & 0x7fff_ffffu32;
107    let xd = x as f64;
108
109    // |x| <= pi/16
110    if x_abs <= 0x3e49_0fdbu32 {
111        // |x| < 0.000443633
112        if x_abs < 0x39e8_9769u32 {
113            if x_abs == 0u32 {
114                // For signed zeros.
115                return if x.is_sign_negative() {
116                    f32::NEG_INFINITY
117                } else {
118                    f32::INFINITY
119                };
120            }
121            let dx = x as f64;
122            let c_term = 1. / dx;
123            let x2 = dx * dx;
124            // Maclaurin series
125            // 1/x + x/6 + (7 x^3)/360 + (31 x^5)/15120 + O(x^7)
126            use crate::polyeval::d_polyeval3;
127            let p = d_polyeval3(
128                x2,
129                f64::from_bits(0x3fc5555555555555),
130                f64::from_bits(0x3f93e93e93e93e94),
131                f64::from_bits(0x3f60b2463814bc5f),
132            );
133            return f64::mul_add(dx, p, c_term) as f32;
134        }
135
136        let xsqr = xd * xd;
137
138        /*
139        Generated by Sollya:
140        f = 1 / sin(x) - 1/x;
141
142        d = [0.000443633; pi/16];
143        pf = fpminimax(f, [|1, 3, 5, 7, 9|], [|D...|], d, relative, floating);
144
145        See ./notes/cscf.sollya
146         */
147        use crate::polyeval::d_polyeval5;
148        let p = d_polyeval5(
149            xsqr,
150            f64::from_bits(0x3fc5555555555562),
151            f64::from_bits(0x3f93e93e93e730a3),
152            f64::from_bits(0x3f60cbb77382ae6f),
153            f64::from_bits(0x3f2b85bfd4188934),
154            f64::from_bits(0x3ef697a32ebe822d),
155        );
156        return f64::mul_add(xd, p, 1. / xd) as f32;
157    }
158
159    if x_abs >= 0x7f80_0000u32 {
160        return x + f32::NAN;
161    }
162
163    // Formula:
164    //   sin(x) = sin((k + y)*pi/32)
165    //          = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32)
166    // The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..31 are precomputed
167    // and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are
168    // computed using degree-7 and degree-6 minimax polynomials generated by
169    // Sollya respectively.
170    use crate::sin_cosf::sincosf_eval::sincosf_eval_fma;
171    let rs = sincosf_eval_fma(xd, x_abs);
172    (1. / f64::mul_add(
173        rs.sin_y,
174        rs.cos_k,
175        f64::mul_add(rs.cosm1_y, rs.sin_k, rs.sin_k),
176    )) as f32
177}
178
179/// Cosecant ( 1 / sin(x) )
180///
181/// ULP 0.5
182#[inline]
183pub fn f_cscf(x: f32) -> f32 {
184    #[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))]
185    {
186        cscf_gen_impl(x)
187    }
188    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
189    {
190        use std::sync::OnceLock;
191        static EXECUTOR: OnceLock<unsafe fn(f32) -> f32> = OnceLock::new();
192        let q = EXECUTOR.get_or_init(|| {
193            if std::arch::is_x86_feature_detected!("avx")
194                && std::arch::is_x86_feature_detected!("fma")
195            {
196                cscf_fma_impl
197            } else {
198                cscf_gen_impl
199            }
200        });
201        unsafe { q(x) }
202    }
203}
204
205#[cfg(test)]
206mod tests {
207    use super::*;
208
209    #[test]
210    fn f_cscf_test() {
211        assert_eq!(f_cscf(0.04915107), 20.353632);
212        assert_eq!(f_cscf(0.5), 2.0858297);
213        assert_eq!(f_cscf(0.07), 14.297387);
214        assert_eq!(f_cscf(3.6171106e-5), 27646.375);
215        assert_eq!(f_cscf(-5.535772e-10), -1806432800.0);
216        assert_eq!(f_cscf(0.0), f32::INFINITY);
217        assert_eq!(f_cscf(-0.0), f32::NEG_INFINITY);
218        assert_eq!(f_cscf(-1.0854926e-19), -9.2124077e18);
219    }
220}