xref: /freebsd/contrib/arm-optimized-routines/pl/math/cbrt_2u.c (revision a03411e84728e9b267056fd31c7d1d9d1dc1b01e)
1 /*
2  * Double-precision cbrt(x) function.
3  *
4  * Copyright (c) 2022-2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "math_config.h"
9 #include "pl_sig.h"
10 #include "pl_test.h"
11 
12 PL_SIG (S, D, 1, cbrt, -10.0, 10.0)
13 
14 #define AbsMask 0x7fffffffffffffff
15 #define TwoThirds 0x1.5555555555555p-1
16 
17 #define C(i) __cbrt_data.poly[i]
18 #define T(i) __cbrt_data.table[i]
19 
20 /* Approximation for double-precision cbrt(x), using low-order polynomial and
21    two Newton iterations. Greatest observed error is 1.79 ULP. Errors repeat
22    according to the exponent, for instance an error observed for double value
23    m * 2^e will be observed for any input m * 2^(e + 3*i), where i is an
24    integer.
25    cbrt(0x1.fffff403f0bc6p+1) got 0x1.965fe72821e9bp+0
26 			     want 0x1.965fe72821e99p+0.  */
27 double
28 cbrt (double x)
29 {
30   uint64_t ix = asuint64 (x);
31   uint64_t iax = ix & AbsMask;
32   uint64_t sign = ix & ~AbsMask;
33 
34   if (unlikely (iax == 0 || iax == 0x7f80000000000000))
35     return x;
36 
37   /* |x| = m * 2^e, where m is in [0.5, 1.0].
38      We can easily decompose x into m and e using frexp.  */
39   int e;
40   double m = frexp (asdouble (iax), &e);
41 
42   /* Calculate rough approximation for cbrt(m) in [0.5, 1.0], starting point for
43      Newton iterations.  */
44   double p_01 = fma (C (1), m, C (0));
45   double p_23 = fma (C (3), m, C (2));
46   double p = fma (p_23, m * m, p_01);
47 
48   /* Two iterations of Newton's method for iteratively approximating cbrt.  */
49   double m_by_3 = m / 3;
50   double a = fma (TwoThirds, p, m_by_3 / (p * p));
51   a = fma (TwoThirds, a, m_by_3 / (a * a));
52 
53   /* Assemble the result by the following:
54 
55      cbrt(x) = cbrt(m) * 2 ^ (e / 3).
56 
57      Let t = (2 ^ (e / 3)) / (2 ^ round(e / 3)).
58 
59      Then we know t = 2 ^ (i / 3), where i is the remainder from e / 3.
60      i is an integer in [-2, 2], so t can be looked up in the table T.
61      Hence the result is assembled as:
62 
63      cbrt(x) = cbrt(m) * t * 2 ^ round(e / 3) * sign.
64      Which can be done easily using ldexp.  */
65   return asdouble (asuint64 (ldexp (a * T (2 + e % 3), e / 3)) | sign);
66 }
67 
68 PL_TEST_ULP (cbrt, 1.30)
69 PL_TEST_INTERVAL (cbrt, 0, inf, 1000000)
70 PL_TEST_INTERVAL (cbrt, -0, -inf, 1000000)
71