1/*
2 * Single-precision vector tanh(x) function.
3 *
4 * Copyright (c) 2022-2023, Arm Limited.
5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6 */
7
8#include "v_math.h"
9#include "pl_sig.h"
10#include "pl_test.h"
11
12#include "v_expm1f_inline.h"
13
14static const struct data
15{
16  struct v_expm1f_data expm1f_consts;
17  uint32x4_t boring_bound, large_bound, onef;
18} data = {
19  .expm1f_consts = V_EXPM1F_DATA,
20  /* 0x1.205966p+3, above which tanhf rounds to 1 (or -1 for  negative).  */
21  .boring_bound = V4 (0x41102cb3),
22  .large_bound = V4 (0x7f800000),
23  .onef = V4 (0x3f800000),
24};
25
26static float32x4_t NOINLINE VPCS_ATTR
27special_case (float32x4_t x, float32x4_t y, uint32x4_t special)
28{
29  return v_call_f32 (tanhf, x, y, special);
30}
31
32/* Approximation for single-precision vector tanh(x), using a simplified
33   version of expm1f. The maximum error is 2.58 ULP:
34   _ZGVnN4v_tanhf (0x1.fa5eep-5) got 0x1.f9ba02p-5
35				want 0x1.f9ba08p-5.  */
36float32x4_t VPCS_ATTR V_NAME_F1 (tanh) (float32x4_t x)
37{
38  const struct data *d = ptr_barrier (&data);
39
40  uint32x4_t ix = vreinterpretq_u32_f32 (x);
41  float32x4_t ax = vabsq_f32 (x);
42  uint32x4_t iax = vreinterpretq_u32_f32 (ax);
43  uint32x4_t sign = veorq_u32 (ix, iax);
44  uint32x4_t is_boring = vcgtq_u32 (iax, d->boring_bound);
45  float32x4_t boring = vreinterpretq_f32_u32 (vorrq_u32 (sign, d->onef));
46
47#if WANT_SIMD_EXCEPT
48  /* If fp exceptions are to be triggered properly, set all special and boring
49     lanes to 0, which will trigger no exceptions, and fix them up later.  */
50  uint32x4_t special = vorrq_u32 (vcgtq_u32 (iax, d->large_bound),
51				  vcltq_u32 (iax, v_u32 (0x34000000)));
52  x = v_zerofy_f32 (x, is_boring);
53  if (unlikely (v_any_u32 (special)))
54    x = v_zerofy_f32 (x, special);
55#else
56  uint32x4_t special = vcgtq_u32 (iax, d->large_bound);
57#endif
58
59  /* tanh(x) = (e^2x - 1) / (e^2x + 1).  */
60  float32x4_t q = expm1f_inline (vmulq_n_f32 (x, 2), &d->expm1f_consts);
61  float32x4_t y = vdivq_f32 (q, vaddq_f32 (q, v_f32 (2.0)));
62  if (unlikely (v_any_u32 (special)))
63    return special_case (vreinterpretq_f32_u32 (ix),
64			 vbslq_f32 (is_boring, boring, y), special);
65  return vbslq_f32 (is_boring, boring, y);
66}
67
68PL_SIG (V, F, 1, tanh, -10.0, 10.0)
69PL_TEST_ULP (V_NAME_F1 (tanh), 2.09)
70PL_TEST_EXPECT_FENV (V_NAME_F1 (tanh), WANT_SIMD_EXCEPT)
71PL_TEST_SYM_INTERVAL (V_NAME_F1 (tanh), 0, 0x1p-23, 1000)
72PL_TEST_SYM_INTERVAL (V_NAME_F1 (tanh), 0x1p-23, 0x1.205966p+3, 100000)
73PL_TEST_SYM_INTERVAL (V_NAME_F1 (tanh), 0x1.205966p+3, inf, 100)
74