| 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 |
*/ |
| 29 |
use crate::common::{f_fmla, is_integerf, is_odd_integerf}; |
| 30 |
use crate::polyeval::f_polyeval5; |
| 31 |
use crate::sin_cosf::sincosf_eval::sincospif_eval; |
| 32 |
|
| 33 |
#[inline(always)] |
| 34 |
fn sincospif_gen_impl(x: f32) -> (f32, f32) { |
| 35 |
let x_abs = x.to_bits() & 0x7fff_ffffu32; |
| 36 |
let xd = x as f64; |
| 37 |
|
| 38 |
// |x| <= 1/16 |
| 39 |
if x_abs <= 0x3d80_0000u32 { |
| 40 |
// |x| < 0.00000009546391 |
| 41 |
if x_abs < 0x38a2_f984u32 { |
| 42 |
const PI: f64 = f64::from_bits(0x400921fb54442d18); |
| 43 |
const MPI_E3_OVER_6: f64 = f64::from_bits(0xc014abbce625be53); |
| 44 |
|
| 45 |
// Small values approximated with Taylor poly for sine |
| 46 |
// x = pi * x - pi^3*x^3/6 |
| 47 |
let x2 = xd * xd; |
| 48 |
let p = f_fmla(x2, MPI_E3_OVER_6, PI); |
| 49 |
let sf = (xd * p) as f32; |
| 50 |
#[cfg(any( |
| 51 |
all( |
| 52 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 53 |
target_feature = "fma" |
| 54 |
), |
| 55 |
target_arch = "aarch64" |
| 56 |
))] |
| 57 |
{ |
| 58 |
use crate::common::f_fmlaf; |
| 59 |
let cs = f_fmlaf(x, f32::from_bits(0xb3000000), 1.); |
| 60 |
return (sf, cs); |
| 61 |
} |
| 62 |
#[cfg(not(any( |
| 63 |
all( |
| 64 |
any(target_arch = "x86", target_arch = "x86_64"), |
| 65 |
target_feature = "fma" |
| 66 |
), |
| 67 |
target_arch = "aarch64" |
| 68 |
)))] |
| 69 |
{ |
| 70 |
let cs = f_fmla(xd, f64::from_bits(0xbe60000000000000), 1.) as f32; |
| 71 |
return (sf, cs); |
| 72 |
} |
| 73 |
} |
| 74 |
|
| 75 |
// Cos(x*PI) |
| 76 |
// Generated poly by Sollya: |
| 77 |
// d = [0, 1/16]; |
| 78 |
// f_cos = cos(y*pi); |
| 79 |
// Q = fpminimax(f_cos, [|0, 2, 4, 6, 8|], [|D...|], d, relative, floating); |
| 80 |
// |
| 81 |
// See ./notes/cospif.sollya |
| 82 |
|
| 83 |
let x2 = xd * xd; |
| 84 |
let cs = f_polyeval5( |
| 85 |
x2, |
| 86 |
f64::from_bits(0x3ff0000000000000), |
| 87 |
f64::from_bits(0xc013bd3cc9be43f7), |
| 88 |
f64::from_bits(0x40103c1f08091fe0), |
| 89 |
f64::from_bits(0xbff55d3ba3d94835), |
| 90 |
f64::from_bits(0x3fce173c2a00e74e), |
| 91 |
) as f32; |
| 92 |
/* |
| 93 |
Sin(x*PI) |
| 94 |
Generated by Sollya: |
| 95 |
d = [0, 1/16]; |
| 96 |
f_sin = sin(y*pi)/y; |
| 97 |
Q = fpminimax(sin(y*pi)/y, [|0, 2, 4, 6, 8|], [|D...|], d, relative, floating); |
| 98 |
|
| 99 |
See ./notes/sinpif.sollya |
| 100 |
*/ |
| 101 |
let p = f_polyeval5( |
| 102 |
x2, |
| 103 |
f64::from_bits(0x400921fb54442d18), |
| 104 |
f64::from_bits(0xc014abbce625bbf2), |
| 105 |
f64::from_bits(0x400466bc675e116a), |
| 106 |
f64::from_bits(0xbfe32d2c0b62d41c), |
| 107 |
f64::from_bits(0x3fb501ec4497cb7d), |
| 108 |
); |
| 109 |
let sf = (xd * p) as f32; |
| 110 |
|
| 111 |
return (sf, cs); |
| 112 |
} |
| 113 |
|
| 114 |
// Numbers greater or equal to 2^23 are always integers or NaN |
| 115 |
if x_abs >= 0x4b00_0000u32 || is_integerf(x) { |
| 116 |
if x_abs >= 0x7f80_0000u32 { |
| 117 |
return (x + f32::NAN, x + f32::NAN); |
| 118 |
} |
| 119 |
static SF: [f32; 2] = [0., -0.]; |
| 120 |
let sf = SF[x.is_sign_negative() as usize]; |
| 121 |
if x_abs < 0x4b80_0000u32 { |
| 122 |
static CF: [f32; 2] = [1., -1.]; |
| 123 |
let cs = CF[is_odd_integerf(x) as usize]; |
| 124 |
return (sf, cs); |
| 125 |
} |
| 126 |
return (sf, 1.); |
| 127 |
} |
| 128 |
|
| 129 |
// Formula: |
| 130 |
// cos(x) = cos((k + y)*pi/32) |
| 131 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 132 |
// The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
| 133 |
// and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
| 134 |
// computed using degree-7 and degree-6 minimax polynomials generated by |
| 135 |
// Sollya respectively. |
| 136 |
// Combine the results with the sine of sum formula: |
| 137 |
// cos(x) = cos((k + y)*pi/32) |
| 138 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 139 |
// = cosm1_y * cos_k + sin_y * sin_k |
| 140 |
// = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
| 141 |
|
| 142 |
// sin(x) = sin((k + y)*pi/32) |
| 143 |
// = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) |
| 144 |
|
| 145 |
let rs = sincospif_eval(xd); |
| 146 |
let cs = f_fmla(rs.sin_y, -rs.sin_k, f_fmla(rs.cosm1_y, rs.cos_k, rs.cos_k)) as f32; |
| 147 |
let sf = f_fmla(rs.sin_y, rs.cos_k, f_fmla(rs.cosm1_y, rs.sin_k, rs.sin_k)) as f32; |
| 148 |
(sf, cs) |
| 149 |
} |
| 150 |
|
| 151 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 152 |
#[target_feature(enable = "avx", enable = "fma")] |
| 153 |
unsafe fn sincospif_fma_impl(x: f32) -> (f32, f32) { |
| 154 |
let x_abs = x.to_bits() & 0x7fff_ffffu32; |
| 155 |
let xd = x as f64; |
| 156 |
|
| 157 |
// |x| <= 1/16 |
| 158 |
if x_abs <= 0x3d80_0000u32 { |
| 159 |
// |x| < 0.00000009546391 |
| 160 |
if x_abs < 0x38a2_f984u32 { |
| 161 |
const PI: f64 = f64::from_bits(0x400921fb54442d18); |
| 162 |
const MPI_E3_OVER_6: f64 = f64::from_bits(0xc014abbce625be53); |
| 163 |
|
| 164 |
// Small values approximated with Taylor poly for sine |
| 165 |
// x = pi * x - pi^3*x^3/6 |
| 166 |
let x2 = xd * xd; |
| 167 |
let p = f64::mul_add(x2, MPI_E3_OVER_6, PI); |
| 168 |
let sf = (xd * p) as f32; |
| 169 |
let cs = f32::mul_add(x, f32::from_bits(0xb3000000), 1.); |
| 170 |
return (sf, cs); |
| 171 |
} |
| 172 |
|
| 173 |
use crate::polyeval::d_polyeval5; |
| 174 |
|
| 175 |
// Cos(x*PI) |
| 176 |
// Generated poly by Sollya: |
| 177 |
// d = [0, 1/16]; |
| 178 |
// f_cos = cos(y*pi); |
| 179 |
// Q = fpminimax(f_cos, [|0, 2, 4, 6, 8|], [|D...|], d, relative, floating); |
| 180 |
// |
| 181 |
// See ./notes/cospif.sollya |
| 182 |
|
| 183 |
let x2 = xd * xd; |
| 184 |
let cs = d_polyeval5( |
| 185 |
x2, |
| 186 |
f64::from_bits(0x3ff0000000000000), |
| 187 |
f64::from_bits(0xc013bd3cc9be43f7), |
| 188 |
f64::from_bits(0x40103c1f08091fe0), |
| 189 |
f64::from_bits(0xbff55d3ba3d94835), |
| 190 |
f64::from_bits(0x3fce173c2a00e74e), |
| 191 |
) as f32; |
| 192 |
/* |
| 193 |
Sin(x*PI) |
| 194 |
Generated by Sollya: |
| 195 |
d = [0, 1/16]; |
| 196 |
f_sin = sin(y*pi)/y; |
| 197 |
Q = fpminimax(sin(y*pi)/y, [|0, 2, 4, 6, 8|], [|D...|], d, relative, floating); |
| 198 |
|
| 199 |
See ./notes/sinpif.sollya |
| 200 |
*/ |
| 201 |
let p = d_polyeval5( |
| 202 |
x2, |
| 203 |
f64::from_bits(0x400921fb54442d18), |
| 204 |
f64::from_bits(0xc014abbce625bbf2), |
| 205 |
f64::from_bits(0x400466bc675e116a), |
| 206 |
f64::from_bits(0xbfe32d2c0b62d41c), |
| 207 |
f64::from_bits(0x3fb501ec4497cb7d), |
| 208 |
); |
| 209 |
let sf = (xd * p) as f32; |
| 210 |
|
| 211 |
return (sf, cs); |
| 212 |
} |
| 213 |
|
| 214 |
// Numbers greater or equal to 2^23 are always integers or NaN |
| 215 |
if x_abs >= 0x4b00_0000u32 || x.round_ties_even() == x { |
| 216 |
if x_abs >= 0x7f80_0000u32 { |
| 217 |
return (x + f32::NAN, x + f32::NAN); |
| 218 |
} |
| 219 |
static SF: [f32; 2] = [0., -0.]; |
| 220 |
let sf = SF[x.is_sign_negative() as usize]; |
| 221 |
if x_abs < 0x4b80_0000u32 { |
| 222 |
static CF: [f32; 2] = [1., -1.]; |
| 223 |
let is_odd_integer = unsafe { (x.to_int_unchecked::<i32>() & 1) != 0 }; |
| 224 |
let cs = CF[is_odd_integer as usize]; |
| 225 |
return (sf, cs); |
| 226 |
} |
| 227 |
return (sf, 1.); |
| 228 |
} |
| 229 |
|
| 230 |
// Formula: |
| 231 |
// cos(x) = cos((k + y)*pi/32) |
| 232 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 233 |
// The values of sin(k*pi/32) and cos(k*pi/32) for k = 0..63 are precomputed |
| 234 |
// and stored using a vector of 32 doubles. Sin(y*pi/32) and cos(y*pi/32) are |
| 235 |
// computed using degree-7 and degree-6 minimax polynomials generated by |
| 236 |
// Sollya respectively. |
| 237 |
// Combine the results with the sine of sum formula: |
| 238 |
// cos(x) = cos((k + y)*pi/32) |
| 239 |
// = cos(y*pi/32) * cos(k*pi/32) - sin(y*pi/32) * sin(k*pi/32) |
| 240 |
// = cosm1_y * cos_k + sin_y * sin_k |
| 241 |
// = (cosm1_y * cos_k + cos_k) + sin_y * sin_k |
| 242 |
|
| 243 |
// sin(x) = sin((k + y)*pi/32) |
| 244 |
// = sin(y*pi/32) * cos(k*pi/32) + cos(y*pi/32) * sin(k*pi/32) |
| 245 |
use crate::sin_cosf::sincosf_eval::sincospif_eval_fma; |
| 246 |
let rs = sincospif_eval_fma(xd); |
| 247 |
let cs = f64::mul_add( |
| 248 |
rs.sin_y, |
| 249 |
-rs.sin_k, |
| 250 |
f64::mul_add(rs.cosm1_y, rs.cos_k, rs.cos_k), |
| 251 |
) as f32; |
| 252 |
let sf = f64::mul_add( |
| 253 |
rs.sin_y, |
| 254 |
rs.cos_k, |
| 255 |
f64::mul_add(rs.cosm1_y, rs.sin_k, rs.sin_k), |
| 256 |
) as f32; |
| 257 |
(sf, cs) |
| 258 |
} |
| 259 |
|
| 260 |
/// Computes sin(x) and cos(x) at the same time |
| 261 |
/// |
| 262 |
/// Max ULP 0.5 |
| 263 |
#[inline] |
| 264 |
pub fn f_sincospif(x: f32) -> (f32, f32) { |
| 265 |
#[cfg(not(any(target_arch = "x86", target_arch = "x86_64")))] |
| 266 |
{ |
| 267 |
sincospif_gen_impl(x) |
| 268 |
} |
| 269 |
#[cfg(any(target_arch = "x86", target_arch = "x86_64"))] |
| 270 |
{ |
| 271 |
use std::sync::OnceLock; |
| 272 |
static EXECUTOR: OnceLock<unsafe fn(f32) -> (f32, f32)> = OnceLock::new(); |
| 273 |
let q = EXECUTOR.get_or_init(|| { |
| 274 |
if std::arch::is_x86_feature_detected!("avx") |
| 275 |
&& std::arch::is_x86_feature_detected!("fma") |
| 276 |
{ |
| 277 |
sincospif_fma_impl |
| 278 |
} else { |
| 279 |
sincospif_gen_impl |
| 280 |
} |
| 281 |
}); |
| 282 |
unsafe { q(x) } |
| 283 |
} |
| 284 |
} |
| 285 |
|
| 286 |
#[cfg(test)] |
| 287 |
mod tests { |
| 288 |
use super::*; |
| 289 |
use crate::{f_cospif, f_sinpif}; |
| 290 |
|
| 291 |
#[test] |
| 292 |
fn test_sincospif() { |
| 293 |
let v0 = f_sincospif(-5.); |
| 294 |
assert_eq!(v0.0, f_sinpif(-5.)); |
| 295 |
assert_eq!(v0.1, f_cospif(-5.)); |
| 296 |
|
| 297 |
let v0 = f_sincospif(-4.); |
| 298 |
assert_eq!(v0.0, f_sinpif(-4.)); |
| 299 |
assert_eq!(v0.1, f_cospif(-4.)); |
| 300 |
|
| 301 |
let v0 = f_sincospif(4.); |
| 302 |
assert_eq!(v0.0, f_sinpif(4.)); |
| 303 |
assert_eq!(v0.1, f_cospif(4.)); |
| 304 |
|
| 305 |
let v0 = f_sincospif(-8489897.0); |
| 306 |
assert_eq!(v0.0, f_sinpif(-8489897.0)); |
| 307 |
assert_eq!(v0.1, f_cospif(-8489897.0)); |
| 308 |
|
| 309 |
let v1 = f_sincospif(3.23); |
| 310 |
assert_eq!(v1.0, f_sinpif(3.23)); |
| 311 |
assert_eq!(v1.1, f_cospif(3.23)); |
| 312 |
} |
| 313 |
} |