xref: /freebsd/contrib/arm-optimized-routines/pl/math/atanh_3u.c (revision 5ca8e32633c4ffbbcd6762e5888b6a4ba0708c6c)
1 /*
2  * Double-precision atanh(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 "poly_scalar_f64.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 #define AbsMask 0x7fffffffffffffff
14 #define Half 0x3fe0000000000000
15 #define One 0x3ff0000000000000
16 #define Ln2Hi 0x1.62e42fefa3800p-1
17 #define Ln2Lo 0x1.ef35793c76730p-45
18 #define OneMHfRt2Top                                                           \
19   0x00095f62 /* top32(asuint64(1)) - top32(asuint64(sqrt(2)/2)).  */
20 #define OneTop12 0x3ff
21 #define HfRt2Top 0x3fe6a09e /* top32(asuint64(sqrt(2)/2)).  */
22 #define BottomMask 0xffffffff
23 
24 static inline double
25 log1p_inline (double x)
26 {
27   /* Helper for calculating log(1 + x) using order-18 polynomial on a reduced
28      interval. Copied from log1p_2u.c, with no special-case handling. See that
29      file for details of the algorithm.  */
30   double m = x + 1;
31   uint64_t mi = asuint64 (m);
32 
33   /* Decompose x + 1 into (f + 1) * 2^k, with k chosen such that f is in
34      [sqrt(2)/2, sqrt(2)].  */
35   uint32_t u = (mi >> 32) + OneMHfRt2Top;
36   int32_t k = (int32_t) (u >> 20) - OneTop12;
37   uint32_t utop = (u & 0x000fffff) + HfRt2Top;
38   uint64_t u_red = ((uint64_t) utop << 32) | (mi & BottomMask);
39   double f = asdouble (u_red) - 1;
40 
41   /* Correction term for round-off in f.  */
42   double cm = (x - (m - 1)) / m;
43 
44   /* Approximate log1p(f) with polynomial.  */
45   double f2 = f * f;
46   double f4 = f2 * f2;
47   double f8 = f4 * f4;
48   double p = fma (
49       f, estrin_18_f64 (f, f2, f4, f8, f8 * f8, __log1p_data.coeffs) * f, f);
50 
51   /* Recombine log1p(x) = k*log2 + log1p(f) + c/m.  */
52   double kd = k;
53   double y = fma (Ln2Lo, kd, cm);
54   return y + fma (Ln2Hi, kd, p);
55 }
56 
57 /* Approximation for double-precision inverse tanh(x), using a simplified
58    version of log1p. Greatest observed error is 3.00 ULP:
59    atanh(0x1.e58f3c108d714p-4) got 0x1.e7da77672a647p-4
60 			      want 0x1.e7da77672a64ap-4.  */
61 double
62 atanh (double x)
63 {
64   uint64_t ix = asuint64 (x);
65   uint64_t sign = ix & ~AbsMask;
66   uint64_t ia = ix & AbsMask;
67 
68   if (unlikely (ia == One))
69     return __math_divzero (sign >> 32);
70 
71   if (unlikely (ia > One))
72     return __math_invalid (x);
73 
74   double halfsign = asdouble (Half | sign);
75   double ax = asdouble (ia);
76   return halfsign * log1p_inline ((2 * ax) / (1 - ax));
77 }
78 
79 PL_SIG (S, D, 1, atanh, -1.0, 1.0)
80 PL_TEST_ULP (atanh, 3.00)
81 PL_TEST_SYM_INTERVAL (atanh, 0, 0x1p-23, 10000)
82 PL_TEST_SYM_INTERVAL (atanh, 0x1p-23, 1, 90000)
83 PL_TEST_SYM_INTERVAL (atanh, 1, inf, 100)
84