1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner * Single-precision vector tanh(x) function.
3*f3087befSAndrew Turner *
4*f3087befSAndrew Turner * Copyright (c) 2022-2024, Arm Limited.
5*f3087befSAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner */
7*f3087befSAndrew Turner
8*f3087befSAndrew Turner #include "v_math.h"
9*f3087befSAndrew Turner #include "test_sig.h"
10*f3087befSAndrew Turner #include "test_defs.h"
11*f3087befSAndrew Turner #include "v_expm1f_inline.h"
12*f3087befSAndrew Turner
13*f3087befSAndrew Turner static const struct data
14*f3087befSAndrew Turner {
15*f3087befSAndrew Turner struct v_expm1f_data expm1f_consts;
16*f3087befSAndrew Turner uint32x4_t boring_bound, large_bound, onef;
17*f3087befSAndrew Turner } data = {
18*f3087befSAndrew Turner .expm1f_consts = V_EXPM1F_DATA,
19*f3087befSAndrew Turner /* 0x1.205966p+3, above which tanhf rounds to 1 (or -1 for negative). */
20*f3087befSAndrew Turner .boring_bound = V4 (0x41102cb3),
21*f3087befSAndrew Turner .large_bound = V4 (0x7f800000),
22*f3087befSAndrew Turner };
23*f3087befSAndrew Turner
24*f3087befSAndrew Turner static float32x4_t NOINLINE VPCS_ATTR
special_case(float32x4_t x,uint32x4_t is_boring,float32x4_t boring,float32x4_t q,uint32x4_t special)25*f3087befSAndrew Turner special_case (float32x4_t x, uint32x4_t is_boring, float32x4_t boring,
26*f3087befSAndrew Turner float32x4_t q, uint32x4_t special)
27*f3087befSAndrew Turner {
28*f3087befSAndrew Turner return v_call_f32 (
29*f3087befSAndrew Turner tanhf, x,
30*f3087befSAndrew Turner vbslq_f32 (is_boring, boring, vdivq_f32 (q, vaddq_f32 (q, v_f32 (2.0)))),
31*f3087befSAndrew Turner special);
32*f3087befSAndrew Turner }
33*f3087befSAndrew Turner
34*f3087befSAndrew Turner /* Approximation for single-precision vector tanh(x), using a simplified
35*f3087befSAndrew Turner version of expm1f. The maximum error is 2.58 ULP:
36*f3087befSAndrew Turner _ZGVnN4v_tanhf (0x1.fa5eep-5) got 0x1.f9ba02p-5
37*f3087befSAndrew Turner want 0x1.f9ba08p-5. */
V_NAME_F1(tanh)38*f3087befSAndrew Turner float32x4_t VPCS_ATTR NOINLINE V_NAME_F1 (tanh) (float32x4_t x)
39*f3087befSAndrew Turner {
40*f3087befSAndrew Turner const struct data *d = ptr_barrier (&data);
41*f3087befSAndrew Turner
42*f3087befSAndrew Turner uint32x4_t ix = vreinterpretq_u32_f32 (x);
43*f3087befSAndrew Turner float32x4_t ax = vabsq_f32 (x);
44*f3087befSAndrew Turner uint32x4_t iax = vreinterpretq_u32_f32 (ax);
45*f3087befSAndrew Turner uint32x4_t sign = veorq_u32 (ix, iax);
46*f3087befSAndrew Turner uint32x4_t is_boring = vcgtq_u32 (iax, d->boring_bound);
47*f3087befSAndrew Turner /* expm1 exponent bias is 1.0f reinterpreted to int. */
48*f3087befSAndrew Turner float32x4_t boring = vreinterpretq_f32_u32 (vorrq_u32 (
49*f3087befSAndrew Turner sign, vreinterpretq_u32_s32 (d->expm1f_consts.exponent_bias)));
50*f3087befSAndrew Turner
51*f3087befSAndrew Turner #if WANT_SIMD_EXCEPT
52*f3087befSAndrew Turner /* If fp exceptions are to be triggered properly, set all special and boring
53*f3087befSAndrew Turner lanes to 0, which will trigger no exceptions, and fix them up later. */
54*f3087befSAndrew Turner uint32x4_t special = vorrq_u32 (vcgtq_u32 (iax, d->large_bound),
55*f3087befSAndrew Turner vcltq_u32 (iax, v_u32 (0x34000000)));
56*f3087befSAndrew Turner x = v_zerofy_f32 (x, is_boring);
57*f3087befSAndrew Turner if (unlikely (v_any_u32 (special)))
58*f3087befSAndrew Turner x = v_zerofy_f32 (x, special);
59*f3087befSAndrew Turner #else
60*f3087befSAndrew Turner uint32x4_t special = vcgtq_u32 (iax, d->large_bound);
61*f3087befSAndrew Turner #endif
62*f3087befSAndrew Turner
63*f3087befSAndrew Turner /* tanh(x) = (e^2x - 1) / (e^2x + 1). */
64*f3087befSAndrew Turner float32x4_t q = expm1f_inline (vmulq_n_f32 (x, 2), &d->expm1f_consts);
65*f3087befSAndrew Turner
66*f3087befSAndrew Turner if (unlikely (v_any_u32 (special)))
67*f3087befSAndrew Turner return special_case (vreinterpretq_f32_u32 (ix), is_boring, boring, q,
68*f3087befSAndrew Turner special);
69*f3087befSAndrew Turner
70*f3087befSAndrew Turner float32x4_t y = vdivq_f32 (q, vaddq_f32 (q, v_f32 (2.0)));
71*f3087befSAndrew Turner return vbslq_f32 (is_boring, boring, y);
72*f3087befSAndrew Turner }
73*f3087befSAndrew Turner
74*f3087befSAndrew Turner HALF_WIDTH_ALIAS_F1 (tanh)
75*f3087befSAndrew Turner
76*f3087befSAndrew Turner TEST_SIG (V, F, 1, tanh, -10.0, 10.0)
77*f3087befSAndrew Turner TEST_ULP (V_NAME_F1 (tanh), 2.09)
78*f3087befSAndrew Turner TEST_DISABLE_FENV_IF_NOT (V_NAME_F1 (tanh), WANT_SIMD_EXCEPT)
79*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (tanh), 0, 0x1p-23, 1000)
80*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (tanh), 0x1p-23, 0x1.205966p+3, 100000)
81*f3087befSAndrew Turner TEST_SYM_INTERVAL (V_NAME_F1 (tanh), 0x1.205966p+3, inf, 100)
82