xref: /freebsd/contrib/arm-optimized-routines/math/aarch64/experimental/tanhf_2u6.c (revision f3087bef11543b42e0d69b708f367097a4118d24)
1 /*
2  * Single-precision tanh(x) function.
3  *
4  * Copyright (c) 2022-2024, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 #include "math_config.h"
8 #include "test_sig.h"
9 #include "test_defs.h"
10 
11 /* 0x1.205966p+3, above which tanhf rounds to 1 (or -1 for negative).  */
12 #define BoringBound 0x41102cb3
13 #define AbsMask 0x7fffffff
14 #define One 0x3f800000
15 
16 #define Shift (0x1.8p23f)
17 #define InvLn2 (0x1.715476p+0f)
18 #define Ln2hi (0x1.62e4p-1f)
19 #define Ln2lo (0x1.7f7d1cp-20f)
20 
21 #define C(i) __expm1f_poly[i]
22 
23 static inline float
expm1f_inline(float x)24 expm1f_inline (float x)
25 {
26   /* Helper routine for calculating exp(x) - 1.
27      Copied from expm1f_1u6.c, with several simplifications:
28      - No special-case handling for tiny or special values, instead return
29        early from the main routine.
30      - No special handling for large values:
31        - No early return for infinity.
32        - Simpler combination of p and t in final stage of algorithm.
33        - |i| < 27, so can calculate t by simpler shift-and-add, instead of
34 	 ldexpf (same as vector algorithm).  */
35 
36   /* Reduce argument: f in [-ln2/2, ln2/2], i is exact.  */
37   float j = fmaf (InvLn2, x, Shift) - Shift;
38   int32_t i = j;
39   float f = fmaf (j, -Ln2hi, x);
40   f = fmaf (j, -Ln2lo, f);
41 
42   /* Approximate expm1(f) with polynomial P, expm1(f) ~= f + f^2 * P(f).
43      Uses Estrin scheme, where the main expm1f routine uses Horner.  */
44   float f2 = f * f;
45   float p_01 = fmaf (f, C (1), C (0));
46   float p_23 = fmaf (f, C (3), C (2));
47   float p = fmaf (f2, p_23, p_01);
48   p = fmaf (f2 * f2, C (4), p);
49   p = fmaf (f2, p, f);
50 
51   /* t = 2^i.  */
52   float t = asfloat ((uint32_t) (i + 127) << 23);
53   /* expm1(x) ~= p * t + (t - 1).  */
54   return fmaf (p, t, t - 1);
55 }
56 
57 /* Approximation for single-precision tanh(x), using a simplified version of
58    expm1f. The maximum error is 2.58 ULP:
59    tanhf(0x1.fa5eep-5) got 0x1.f9ba02p-5
60 		      want 0x1.f9ba08p-5.  */
61 float
tanhf(float x)62 tanhf (float x)
63 {
64   uint32_t ix = asuint (x);
65   uint32_t iax = ix & AbsMask;
66   uint32_t sign = ix & ~AbsMask;
67 
68   if (unlikely (iax > BoringBound))
69     {
70       if (iax > 0x7f800000)
71 	return __math_invalidf (x);
72       return asfloat (One | sign);
73     }
74 
75   if (unlikely (iax < 0x34000000))
76     return x;
77 
78   /* tanh(x) = (e^2x - 1) / (e^2x + 1).  */
79   float q = expm1f_inline (2 * x);
80   return q / (q + 2);
81 }
82 
83 TEST_SIG (S, F, 1, tanh, -10.0, 10.0)
84 TEST_ULP (tanhf, 2.09)
85 TEST_SYM_INTERVAL (tanhf, 0, 0x1p-23, 1000)
86 TEST_SYM_INTERVAL (tanhf, 0x1p-23, 0x1.205966p+3, 100000)
87 TEST_SYM_INTERVAL (tanhf, 0x1.205966p+3, inf, 100)
88