1 | /* SPDX-License-Identifier: MIT */ |
2 | /* origin: core-math/src/binary64/cbrt/cbrt.c |
3 | * Copyright (c) 2021-2022 Alexei Sibidanov. |
4 | * Ported to Rust in 2025 by Trevor Gross. |
5 | */ |
6 | |
7 | use super::Float; |
8 | use super::support::{FpResult, Round, cold_path}; |
9 | |
10 | /// Compute the cube root of the argument. |
11 | #[cfg_attr (all(test, assert_no_panic), no_panic::no_panic)] |
12 | pub fn cbrt(x: f64) -> f64 { |
13 | cbrt_round(x, Round::Nearest).val |
14 | } |
15 | |
16 | pub fn cbrt_round(x: f64, round: Round) -> FpResult<f64> { |
17 | const ESCALE: [f64; 3] = [ |
18 | 1.0, |
19 | hf64!("0x1.428a2f98d728bp+0" ), /* 2^(1/3) */ |
20 | hf64!("0x1.965fea53d6e3dp+0" ), /* 2^(2/3) */ |
21 | ]; |
22 | |
23 | /* the polynomial c0+c1*x+c2*x^2+c3*x^3 approximates x^(1/3) on [1,2] |
24 | with maximal error < 9.2e-5 (attained at x=2) */ |
25 | const C: [f64; 4] = [ |
26 | hf64!("0x1.1b0babccfef9cp-1" ), |
27 | hf64!("0x1.2c9a3e94d1da5p-1" ), |
28 | hf64!("-0x1.4dc30b1a1ddbap-3" ), |
29 | hf64!("0x1.7a8d3e4ec9b07p-6" ), |
30 | ]; |
31 | |
32 | let u0: f64 = hf64!("0x1.5555555555555p-2" ); |
33 | let u1: f64 = hf64!("0x1.c71c71c71c71cp-3" ); |
34 | |
35 | let rsc = [1.0, -1.0, 0.5, -0.5, 0.25, -0.25]; |
36 | |
37 | let off = [hf64!("0x1p-53" ), 0.0, 0.0, 0.0]; |
38 | |
39 | /* rm=0 for rounding to nearest, and other values for directed roundings */ |
40 | let hx: u64 = x.to_bits(); |
41 | let mut mant: u64 = hx & f64::SIG_MASK; |
42 | let sign: u64 = hx >> 63; |
43 | |
44 | let mut e: u32 = (hx >> f64::SIG_BITS) as u32 & f64::EXP_SAT; |
45 | |
46 | if ((e + 1) & f64::EXP_SAT) < 2 { |
47 | cold_path(); |
48 | |
49 | let ix: u64 = hx & !f64::SIGN_MASK; |
50 | |
51 | /* 0, inf, nan: we return x + x instead of simply x, |
52 | to that for x a signaling NaN, it correctly triggers |
53 | the invalid exception. */ |
54 | if e == f64::EXP_SAT || ix == 0 { |
55 | return FpResult::ok(x + x); |
56 | } |
57 | |
58 | let nz = ix.leading_zeros() - 11; /* subnormal */ |
59 | mant <<= nz; |
60 | mant &= f64::SIG_MASK; |
61 | e = e.wrapping_sub(nz - 1); |
62 | } |
63 | |
64 | e = e.wrapping_add(3072); |
65 | let cvt1: u64 = mant | (0x3ffu64 << 52); |
66 | let mut cvt5: u64 = cvt1; |
67 | |
68 | let et: u32 = e / 3; |
69 | let it: u32 = e % 3; |
70 | |
71 | /* 2^(3k+it) <= x < 2^(3k+it+1), with 0 <= it <= 3 */ |
72 | cvt5 += u64::from(it) << f64::SIG_BITS; |
73 | cvt5 |= sign << 63; |
74 | let zz: f64 = f64::from_bits(cvt5); |
75 | |
76 | /* cbrt(x) = cbrt(zz)*2^(et-1365) where 1 <= zz < 8 */ |
77 | let mut isc: u64 = ESCALE[it as usize].to_bits(); // todo: index |
78 | isc |= sign << 63; |
79 | let cvt2: u64 = isc; |
80 | let z: f64 = f64::from_bits(cvt1); |
81 | |
82 | /* cbrt(zz) = cbrt(z)*isc, where isc encodes 1, 2^(1/3) or 2^(2/3), |
83 | and 1 <= z < 2 */ |
84 | let r: f64 = 1.0 / z; |
85 | let rr: f64 = r * rsc[((it as usize) << 1) | sign as usize]; |
86 | let z2: f64 = z * z; |
87 | let c0: f64 = C[0] + z * C[1]; |
88 | let c2: f64 = C[2] + z * C[3]; |
89 | let mut y: f64 = c0 + z2 * c2; |
90 | let mut y2: f64 = y * y; |
91 | |
92 | /* y is an approximation of z^(1/3) */ |
93 | let mut h: f64 = y2 * (y * r) - 1.0; |
94 | |
95 | /* h determines the error between y and z^(1/3) */ |
96 | y -= (h * y) * (u0 - u1 * h); |
97 | |
98 | /* The correction y -= (h*y)*(u0 - u1*h) corresponds to a cubic variant |
99 | of Newton's method, with the function f(y) = 1-z/y^3. */ |
100 | y *= f64::from_bits(cvt2); |
101 | |
102 | /* Now y is an approximation of zz^(1/3), |
103 | * and rr an approximation of 1/zz. We now perform another iteration of |
104 | * Newton-Raphson, this time with a linear approximation only. */ |
105 | y2 = y * y; |
106 | let mut y2l: f64 = y.fma(y, -y2); |
107 | |
108 | /* y2 + y2l = y^2 exactly */ |
109 | let mut y3: f64 = y2 * y; |
110 | let mut y3l: f64 = y.fma(y2, -y3) + y * y2l; |
111 | |
112 | /* y3 + y3l approximates y^3 with about 106 bits of accuracy */ |
113 | h = ((y3 - zz) + y3l) * rr; |
114 | let mut dy: f64 = h * (y * u0); |
115 | |
116 | /* the approximation of zz^(1/3) is y - dy */ |
117 | let mut y1: f64 = y - dy; |
118 | dy = (y - y1) - dy; |
119 | |
120 | /* the approximation of zz^(1/3) is now y1 + dy, where |dy| < 1/2 ulp(y) |
121 | * (for rounding to nearest) */ |
122 | let mut ady: f64 = dy.abs(); |
123 | |
124 | /* For directed roundings, ady0 is tiny when dy is tiny, or ady0 is near |
125 | * from ulp(1); |
126 | * for rounding to nearest, ady0 is tiny when dy is near from 1/2 ulp(1), |
127 | * or from 3/2 ulp(1). */ |
128 | let mut ady0: f64 = (ady - off[round as usize]).abs(); |
129 | let mut ady1: f64 = (ady - (hf64!("0x1p-52" ) + off[round as usize])).abs(); |
130 | |
131 | if ady0 < hf64!("0x1p-75" ) || ady1 < hf64!("0x1p-75" ) { |
132 | cold_path(); |
133 | |
134 | y2 = y1 * y1; |
135 | y2l = y1.fma(y1, -y2); |
136 | y3 = y2 * y1; |
137 | y3l = y1.fma(y2, -y3) + y1 * y2l; |
138 | h = ((y3 - zz) + y3l) * rr; |
139 | dy = h * (y1 * u0); |
140 | y = y1 - dy; |
141 | dy = (y1 - y) - dy; |
142 | y1 = y; |
143 | ady = dy.abs(); |
144 | ady0 = (ady - off[round as usize]).abs(); |
145 | ady1 = (ady - (hf64!("0x1p-52" ) + off[round as usize])).abs(); |
146 | |
147 | if ady0 < hf64!("0x1p-98" ) || ady1 < hf64!("0x1p-98" ) { |
148 | cold_path(); |
149 | let azz: f64 = zz.abs(); |
150 | |
151 | // ~ 0x1.79d15d0e8d59b80000000000000ffc3dp+0 |
152 | if azz == hf64!("0x1.9b78223aa307cp+1" ) { |
153 | y1 = hf64!("0x1.79d15d0e8d59cp+0" ).copysign(zz); |
154 | } |
155 | |
156 | // ~ 0x1.de87aa837820e80000000000001c0f08p+0 |
157 | if azz == hf64!("0x1.a202bfc89ddffp+2" ) { |
158 | y1 = hf64!("0x1.de87aa837820fp+0" ).copysign(zz); |
159 | } |
160 | |
161 | if round != Round::Nearest { |
162 | let wlist = [ |
163 | (hf64!("0x1.3a9ccd7f022dbp+0" ), hf64!("0x1.1236160ba9b93p+0" )), // ~ 0x1.1236160ba9b930000000000001e7e8fap+0 |
164 | (hf64!("0x1.7845d2faac6fep+0" ), hf64!("0x1.23115e657e49cp+0" )), // ~ 0x1.23115e657e49c0000000000001d7a799p+0 |
165 | (hf64!("0x1.d1ef81cbbbe71p+0" ), hf64!("0x1.388fb44cdcf5ap+0" )), // ~ 0x1.388fb44cdcf5a0000000000002202c55p+0 |
166 | (hf64!("0x1.0a2014f62987cp+1" ), hf64!("0x1.46bcbf47dc1e8p+0" )), // ~ 0x1.46bcbf47dc1e8000000000000303aa2dp+0 |
167 | (hf64!("0x1.fe18a044a5501p+1" ), hf64!("0x1.95decfec9c904p+0" )), // ~ 0x1.95decfec9c9040000000000000159e8ep+0 |
168 | (hf64!("0x1.a6bb8c803147bp+2" ), hf64!("0x1.e05335a6401dep+0" )), // ~ 0x1.e05335a6401de00000000000027ca017p+0 |
169 | (hf64!("0x1.ac8538a031cbdp+2" ), hf64!("0x1.e281d87098de8p+0" )), // ~ 0x1.e281d87098de80000000000000ee9314p+0 |
170 | ]; |
171 | |
172 | for (a, b) in wlist { |
173 | if azz == a { |
174 | let tmp = if round as u64 + sign == 2 { |
175 | hf64!("0x1p-52" ) |
176 | } else { |
177 | 0.0 |
178 | }; |
179 | y1 = (b + tmp).copysign(zz); |
180 | } |
181 | } |
182 | } |
183 | } |
184 | } |
185 | |
186 | let mut cvt3: u64 = y1.to_bits(); |
187 | cvt3 = cvt3.wrapping_add(((et.wrapping_sub(342).wrapping_sub(1023)) as u64) << 52); |
188 | let m0: u64 = cvt3 << 30; |
189 | let m1 = m0 >> 63; |
190 | |
191 | if (m0 ^ m1) <= (1u64 << 30) { |
192 | cold_path(); |
193 | |
194 | let mut cvt4: u64 = y1.to_bits(); |
195 | cvt4 = (cvt4 + (164 << 15)) & 0xffffffffffff0000u64; |
196 | |
197 | if ((f64::from_bits(cvt4) - y1) - dy).abs() < hf64!("0x1p-60" ) || (zz).abs() == 1.0 { |
198 | cvt3 = (cvt3 + (1u64 << 15)) & 0xffffffffffff0000u64; |
199 | } |
200 | } |
201 | |
202 | FpResult::ok(f64::from_bits(cvt3)) |
203 | } |
204 | |
205 | #[cfg (test)] |
206 | mod tests { |
207 | use super::*; |
208 | |
209 | #[test ] |
210 | fn spot_checks() { |
211 | if !cfg!(x86_no_sse) { |
212 | // Exposes a rounding mode problem. Ignored on i586 because of inaccurate FMA. |
213 | assert_biteq!( |
214 | cbrt(f64::from_bits(0xf7f792b28f600000)), |
215 | f64::from_bits(0xd29ce68655d962f3) |
216 | ); |
217 | } |
218 | } |
219 | } |
220 | |