xref: /freebsd/contrib/arm-optimized-routines/pl/math/cbrtf_1u5.c (revision 59c8e88e72633afbc47a4ace0d2170d00d51f7dc)
1 /*
2  * Single-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 "poly_scalar_f32.h"
9 #include "math_config.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 #define AbsMask 0x7fffffff
14 #define SignMask 0x80000000
15 #define TwoThirds 0x1.555556p-1f
16 
17 #define T(i) __cbrtf_data.table[i]
18 
19 /* Approximation for single-precision cbrt(x), using low-order polynomial and
20    one Newton iteration on a reduced interval. Greatest error is 1.5 ULP. This
21    is observed for every value where the mantissa is 0x1.81410e and the exponent
22    is a multiple of 3, for example:
23    cbrtf(0x1.81410ep+30) got 0x1.255d96p+10
24 			want 0x1.255d92p+10.  */
25 float
26 cbrtf (float x)
27 {
28   uint32_t ix = asuint (x);
29   uint32_t iax = ix & AbsMask;
30   uint32_t sign = ix & SignMask;
31 
32   if (unlikely (iax == 0 || iax == 0x7f800000))
33     return x;
34 
35   /* |x| = m * 2^e, where m is in [0.5, 1.0].
36      We can easily decompose x into m and e using frexpf.  */
37   int e;
38   float m = frexpf (asfloat (iax), &e);
39 
40   /* p is a rough approximation for cbrt(m) in [0.5, 1.0]. The better this is,
41      the less accurate the next stage of the algorithm needs to be. An order-4
42      polynomial is enough for one Newton iteration.  */
43   float p = pairwise_poly_3_f32 (m, m * m, __cbrtf_data.poly);
44 
45   /* One iteration of Newton's method for iteratively approximating cbrt.  */
46   float m_by_3 = m / 3;
47   float a = fmaf (TwoThirds, p, m_by_3 / (p * p));
48 
49   /* Assemble the result by the following:
50 
51      cbrt(x) = cbrt(m) * 2 ^ (e / 3).
52 
53      Let t = (2 ^ (e / 3)) / (2 ^ round(e / 3)).
54 
55      Then we know t = 2 ^ (i / 3), where i is the remainder from e / 3.
56      i is an integer in [-2, 2], so t can be looked up in the table T.
57      Hence the result is assembled as:
58 
59      cbrt(x) = cbrt(m) * t * 2 ^ round(e / 3) * sign.
60      Which can be done easily using ldexpf.  */
61   return asfloat (asuint (ldexpf (a * T (2 + e % 3), e / 3)) | sign);
62 }
63 
64 PL_SIG (S, F, 1, cbrt, -10.0, 10.0)
65 PL_TEST_ULP (cbrtf, 1.03)
66 PL_TEST_SYM_INTERVAL (cbrtf, 0, inf, 1000000)
67