xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_tanh_3u.c (revision 5a02ffc32e777041dd2dad4e651ed2a0865a0a5d)
1 /*
2  * Double-precision vector tanh(x) function.
3  * Copyright (c) 2023, Arm Limited.
4  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
5  */
6 
7 #include "v_math.h"
8 #include "poly_advsimd_f64.h"
9 #include "mathlib.h"
10 #include "pl_sig.h"
11 #include "pl_test.h"
12 
13 static const struct data
14 {
15   float64x2_t poly[11];
16   float64x2_t inv_ln2, ln2_hi, ln2_lo, shift;
17   uint64x2_t onef;
18   uint64x2_t thresh, tiny_bound;
19 } data = {
20   /* Generated using Remez, deg=12 in [-log(2)/2, log(2)/2].  */
21   .poly = { V2 (0x1p-1), V2 (0x1.5555555555559p-3), V2 (0x1.555555555554bp-5),
22 	    V2 (0x1.111111110f663p-7), V2 (0x1.6c16c16c1b5f3p-10),
23 	    V2 (0x1.a01a01affa35dp-13), V2 (0x1.a01a018b4ecbbp-16),
24 	    V2 (0x1.71ddf82db5bb4p-19), V2 (0x1.27e517fc0d54bp-22),
25 	    V2 (0x1.af5eedae67435p-26), V2 (0x1.1f143d060a28ap-29), },
26 
27   .inv_ln2 = V2 (0x1.71547652b82fep0),
28   .ln2_hi = V2 (-0x1.62e42fefa39efp-1),
29   .ln2_lo = V2 (-0x1.abc9e3b39803fp-56),
30   .shift = V2 (0x1.8p52),
31 
32   .onef = V2 (0x3ff0000000000000),
33   .tiny_bound = V2 (0x3e40000000000000), /* asuint64 (0x1p-27).  */
34   /* asuint64(0x1.241bf835f9d5fp+4) - asuint64(tiny_bound).  */
35   .thresh = V2 (0x01f241bf835f9d5f),
36 };
37 
38 static inline float64x2_t
expm1_inline(float64x2_t x,const struct data * d)39 expm1_inline (float64x2_t x, const struct data *d)
40 {
41   /* Helper routine for calculating exp(x) - 1. Vector port of the helper from
42      the scalar variant of tanh.  */
43 
44   /* Reduce argument: f in [-ln2/2, ln2/2], i is exact.  */
45   float64x2_t j = vsubq_f64 (vfmaq_f64 (d->shift, d->inv_ln2, x), d->shift);
46   int64x2_t i = vcvtq_s64_f64 (j);
47   float64x2_t f = vfmaq_f64 (x, j, d->ln2_hi);
48   f = vfmaq_f64 (f, j, d->ln2_lo);
49 
50   /* Approximate expm1(f) using polynomial.  */
51   float64x2_t f2 = vmulq_f64 (f, f);
52   float64x2_t f4 = vmulq_f64 (f2, f2);
53   float64x2_t p = vfmaq_f64 (
54       f, f2, v_estrin_10_f64 (f, f2, f4, vmulq_f64 (f4, f4), d->poly));
55 
56   /* t = 2 ^ i.  */
57   float64x2_t t = vreinterpretq_f64_u64 (
58       vaddq_u64 (vreinterpretq_u64_s64 (i << 52), d->onef));
59   /* expm1(x) = p * t + (t - 1).  */
60   return vfmaq_f64 (vsubq_f64 (t, v_f64 (1)), p, t);
61 }
62 
63 static float64x2_t NOINLINE VPCS_ATTR
special_case(float64x2_t x,float64x2_t y,uint64x2_t special)64 special_case (float64x2_t x, float64x2_t y, uint64x2_t special)
65 {
66   return v_call_f64 (tanh, x, y, special);
67 }
68 
69 /* Vector approximation for double-precision tanh(x), using a simplified
70    version of expm1. The greatest observed error is 2.77 ULP:
71    _ZGVnN2v_tanh(-0x1.c4a4ca0f9f3b7p-3) got -0x1.bd6a21a163627p-3
72 				       want -0x1.bd6a21a163624p-3.  */
V_NAME_D1(tanh)73 float64x2_t VPCS_ATTR V_NAME_D1 (tanh) (float64x2_t x)
74 {
75   const struct data *d = ptr_barrier (&data);
76 
77   uint64x2_t ia = vreinterpretq_u64_f64 (vabsq_f64 (x));
78 
79   float64x2_t u = x;
80 
81   /* Trigger special-cases for tiny, boring and infinity/NaN.  */
82   uint64x2_t special = vcgtq_u64 (vsubq_u64 (ia, d->tiny_bound), d->thresh);
83 #if WANT_SIMD_EXCEPT
84   /* To trigger fp exceptions correctly, set special lanes to a neutral value.
85      They will be fixed up later by the special-case handler.  */
86   if (unlikely (v_any_u64 (special)))
87     u = v_zerofy_f64 (u, special);
88 #endif
89 
90   u = vaddq_f64 (u, u);
91 
92   /* tanh(x) = (e^2x - 1) / (e^2x + 1).  */
93   float64x2_t q = expm1_inline (u, d);
94   float64x2_t qp2 = vaddq_f64 (q, v_f64 (2));
95 
96   if (unlikely (v_any_u64 (special)))
97     return special_case (x, vdivq_f64 (q, qp2), special);
98   return vdivq_f64 (q, qp2);
99 }
100 
101 PL_SIG (V, D, 1, tanh, -10.0, 10.0)
102 PL_TEST_ULP (V_NAME_D1 (tanh), 2.27)
103 PL_TEST_EXPECT_FENV (V_NAME_D1 (tanh), WANT_SIMD_EXCEPT)
104 PL_TEST_SYM_INTERVAL (V_NAME_D1 (tanh), 0, 0x1p-27, 5000)
105 PL_TEST_SYM_INTERVAL (V_NAME_D1 (tanh), 0x1p-27, 0x1.241bf835f9d5fp+4, 50000)
106 PL_TEST_SYM_INTERVAL (V_NAME_D1 (tanh), 0x1.241bf835f9d5fp+4, inf, 1000)
107