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