xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/sve/cbrt.c (revision dd21556857e8d40f66bf5ad54754d9d52669ebf7)
1 /*
2  * Double-precision SVE cbrt(x) function.
3  *
4  * Copyright (c) 2023-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "sv_math.h"
9 #include "test_sig.h"
10 #include "test_defs.h"
11 #include "sv_poly_f64.h"
12 
13 const static struct data
14 {
15   float64_t poly[4];
16   float64_t table[5];
17   float64_t one_third, two_thirds, shift;
18   int64_t exp_bias;
19   uint64_t tiny_bound, thresh;
20 } data = {
21   /* Generated with FPMinimax in [0.5, 1].  */
22   .poly = { 0x1.c14e8ee44767p-2, 0x1.dd2d3f99e4c0ep-1, -0x1.08e83026b7e74p-1,
23 	    0x1.2c74eaa3ba428p-3, },
24   /* table[i] = 2^((i - 2) / 3).  */
25   .table = { 0x1.428a2f98d728bp-1, 0x1.965fea53d6e3dp-1, 0x1p0,
26 	     0x1.428a2f98d728bp0, 0x1.965fea53d6e3dp0, },
27   .one_third = 0x1.5555555555555p-2,
28   .two_thirds = 0x1.5555555555555p-1,
29   .shift = 0x1.8p52,
30   .exp_bias = 1022,
31   .tiny_bound = 0x0010000000000000, /* Smallest normal.  */
32   .thresh = 0x7fe0000000000000, /* asuint64 (infinity) - tiny_bound.  */
33 };
34 
35 #define MantissaMask 0x000fffffffffffff
36 #define HalfExp 0x3fe0000000000000
37 
38 static svfloat64_t NOINLINE
39 special_case (svfloat64_t x, svfloat64_t y, svbool_t special)
40 {
41   return sv_call_f64 (cbrt, x, y, special);
42 }
43 
44 static inline svfloat64_t
45 shifted_lookup (const svbool_t pg, const float64_t *table, svint64_t i)
46 {
47   return svld1_gather_index (pg, table, svadd_x (pg, i, 2));
48 }
49 
50 /* Approximation for double-precision vector cbrt(x), using low-order
51    polynomial and two Newton iterations.
52 
53    The vector version of frexp does not handle subnormals
54    correctly. As a result these need to be handled by the scalar
55    fallback, where accuracy may be worse than that of the vector code
56    path.
57 
58    Greatest observed error in the normal range is 1.79 ULP. Errors repeat
59    according to the exponent, for instance an error observed for double value m
60    * 2^e will be observed for any input m * 2^(e + 3*i), where i is an integer.
61    _ZGVsMxv_cbrt (0x0.3fffb8d4413f3p-1022) got 0x1.965f53b0e5d97p-342
62 					  want 0x1.965f53b0e5d95p-342.  */
63 svfloat64_t SV_NAME_D1 (cbrt) (svfloat64_t x, const svbool_t pg)
64 {
65   const struct data *d = ptr_barrier (&data);
66 
67   svfloat64_t ax = svabs_x (pg, x);
68   svuint64_t iax = svreinterpret_u64 (ax);
69   svuint64_t sign = sveor_x (pg, svreinterpret_u64 (x), iax);
70 
71   /* Subnormal, +/-0 and special values.  */
72   svbool_t special = svcmpge (pg, svsub_x (pg, iax, d->tiny_bound), d->thresh);
73 
74   /* Decompose |x| into m * 2^e, where m is in [0.5, 1.0]. This is a vector
75      version of frexp, which gets subnormal values wrong - these have to be
76      special-cased as a result.  */
77   svfloat64_t m = svreinterpret_f64 (svorr_x (
78       pg, svand_x (pg, svreinterpret_u64 (x), MantissaMask), HalfExp));
79   svint64_t e
80       = svsub_x (pg, svreinterpret_s64 (svlsr_x (pg, iax, 52)), d->exp_bias);
81 
82   /* Calculate rough approximation for cbrt(m) in [0.5, 1.0], starting point
83      for Newton iterations.  */
84   svfloat64_t p
85       = sv_pairwise_poly_3_f64_x (pg, m, svmul_x (pg, m, m), d->poly);
86 
87   /* Two iterations of Newton's method for iteratively approximating cbrt.  */
88   svfloat64_t m_by_3 = svmul_x (pg, m, d->one_third);
89   svfloat64_t a = svmla_x (pg, svdiv_x (pg, m_by_3, svmul_x (pg, p, p)), p,
90 			   d->two_thirds);
91   a = svmla_x (pg, svdiv_x (pg, m_by_3, svmul_x (pg, a, a)), a, d->two_thirds);
92 
93   /* Assemble the result by the following:
94 
95      cbrt(x) = cbrt(m) * 2 ^ (e / 3).
96 
97      We can get 2 ^ round(e / 3) using ldexp and integer divide, but since e is
98      not necessarily a multiple of 3 we lose some information.
99 
100      Let q = 2 ^ round(e / 3), then t = 2 ^ (e / 3) / q.
101 
102      Then we know t = 2 ^ (i / 3), where i is the remainder from e / 3, which
103      is an integer in [-2, 2], and can be looked up in the table T. Hence the
104      result is assembled as:
105 
106      cbrt(x) = cbrt(m) * t * 2 ^ round(e / 3) * sign.  */
107   svfloat64_t eb3f = svmul_x (pg, svcvt_f64_x (pg, e), d->one_third);
108   svint64_t ey = svcvt_s64_x (pg, eb3f);
109   svint64_t em3 = svmls_x (pg, e, ey, 3);
110 
111   svfloat64_t my = shifted_lookup (pg, d->table, em3);
112   my = svmul_x (pg, my, a);
113 
114   /* Vector version of ldexp.  */
115   svfloat64_t y = svscale_x (pg, my, ey);
116 
117   if (unlikely (svptest_any (pg, special)))
118     return special_case (
119 	x, svreinterpret_f64 (svorr_x (pg, svreinterpret_u64 (y), sign)),
120 	special);
121 
122   /* Copy sign.  */
123   return svreinterpret_f64 (svorr_x (pg, svreinterpret_u64 (y), sign));
124 }
125 
126 /* Worse-case ULP error assumes that scalar fallback is GLIBC 2.40 cbrt, which
127    has ULP error of 3.67 at 0x1.7a337e1ba1ec2p-257 [1]. Largest observed error
128    in the vector path is 1.79 ULP.
129    [1] Innocente, V., & Zimmermann, P. (2024). Accuracy of Mathematical
130    Functions in Single, Double, Double Extended, and Quadruple Precision.  */
131 TEST_SIG (SV, D, 1, cbrt, -10.0, 10.0)
132 TEST_ULP (SV_NAME_D1 (cbrt), 3.17)
133 TEST_DISABLE_FENV (SV_NAME_D1 (cbrt))
134 TEST_SYM_INTERVAL (SV_NAME_D1 (cbrt), 0, inf, 1000000)
135 CLOSE_SVE_ATTR
136