1*412f47f9SXin Li /*
2*412f47f9SXin Li * Single-precision SVE tanh(x) function.
3*412f47f9SXin Li *
4*412f47f9SXin Li * Copyright (c) 2023, Arm Limited.
5*412f47f9SXin Li * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*412f47f9SXin Li */
7*412f47f9SXin Li
8*412f47f9SXin Li #include "sv_math.h"
9*412f47f9SXin Li #include "pl_sig.h"
10*412f47f9SXin Li #include "pl_test.h"
11*412f47f9SXin Li
12*412f47f9SXin Li #include "sv_expm1f_inline.h"
13*412f47f9SXin Li
14*412f47f9SXin Li static const struct data
15*412f47f9SXin Li {
16*412f47f9SXin Li struct sv_expm1f_data expm1f_consts;
17*412f47f9SXin Li uint32_t boring_bound, onef;
18*412f47f9SXin Li } data = {
19*412f47f9SXin Li .expm1f_consts = SV_EXPM1F_DATA,
20*412f47f9SXin Li /* 0x1.205966p+3, above which tanhf rounds to 1 (or -1 for negative). */
21*412f47f9SXin Li .boring_bound = 0x41102cb3,
22*412f47f9SXin Li .onef = 0x3f800000,
23*412f47f9SXin Li };
24*412f47f9SXin Li
25*412f47f9SXin Li static svfloat32_t NOINLINE
special_case(svfloat32_t x,svfloat32_t y,svbool_t special)26*412f47f9SXin Li special_case (svfloat32_t x, svfloat32_t y, svbool_t special)
27*412f47f9SXin Li {
28*412f47f9SXin Li return sv_call_f32 (tanhf, x, y, special);
29*412f47f9SXin Li }
30*412f47f9SXin Li
31*412f47f9SXin Li /* Approximation for single-precision SVE tanh(x), using a simplified
32*412f47f9SXin Li version of expm1f. The maximum error is 2.57 ULP:
33*412f47f9SXin Li _ZGVsMxv_tanhf (0x1.fc1832p-5) got 0x1.fb71a4p-5
34*412f47f9SXin Li want 0x1.fb71aap-5. */
SV_NAME_F1(tanh)35*412f47f9SXin Li svfloat32_t SV_NAME_F1 (tanh) (svfloat32_t x, const svbool_t pg)
36*412f47f9SXin Li {
37*412f47f9SXin Li const struct data *d = ptr_barrier (&data);
38*412f47f9SXin Li
39*412f47f9SXin Li svfloat32_t ax = svabs_x (pg, x);
40*412f47f9SXin Li svuint32_t iax = svreinterpret_u32 (ax);
41*412f47f9SXin Li svuint32_t sign = sveor_x (pg, svreinterpret_u32 (x), iax);
42*412f47f9SXin Li svbool_t is_boring = svcmpgt (pg, iax, d->boring_bound);
43*412f47f9SXin Li svfloat32_t boring = svreinterpret_f32 (svorr_x (pg, sign, d->onef));
44*412f47f9SXin Li
45*412f47f9SXin Li svbool_t special = svcmpgt (pg, iax, 0x7f800000);
46*412f47f9SXin Li
47*412f47f9SXin Li /* tanh(x) = (e^2x - 1) / (e^2x + 1). */
48*412f47f9SXin Li svfloat32_t q = expm1f_inline (svmul_x (pg, x, 2.0), pg, &d->expm1f_consts);
49*412f47f9SXin Li svfloat32_t y = svdiv_x (pg, q, svadd_x (pg, q, 2.0));
50*412f47f9SXin Li if (unlikely (svptest_any (pg, special)))
51*412f47f9SXin Li return special_case (x, svsel_f32 (is_boring, boring, y), special);
52*412f47f9SXin Li return svsel_f32 (is_boring, boring, y);
53*412f47f9SXin Li }
54*412f47f9SXin Li
55*412f47f9SXin Li PL_SIG (SV, F, 1, tanh, -10.0, 10.0)
56*412f47f9SXin Li PL_TEST_ULP (SV_NAME_F1 (tanh), 2.07)
57*412f47f9SXin Li PL_TEST_SYM_INTERVAL (SV_NAME_F1 (tanh), 0, 0x1p-23, 1000)
58*412f47f9SXin Li PL_TEST_SYM_INTERVAL (SV_NAME_F1 (tanh), 0x1p-23, 0x1.205966p+3, 100000)
59*412f47f9SXin Li PL_TEST_SYM_INTERVAL (SV_NAME_F1 (tanh), 0x1.205966p+3, inf, 100)
60