1*412f47f9SXin Li /*
2*412f47f9SXin Li * Single-precision SVE sinh(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 halff, large_bound;
18*412f47f9SXin Li } data = {
19*412f47f9SXin Li .expm1f_consts = SV_EXPM1F_DATA,
20*412f47f9SXin Li .halff = 0x3f000000,
21*412f47f9SXin Li /* 0x1.61814ep+6, above which expm1f helper overflows. */
22*412f47f9SXin Li .large_bound = 0x42b0c0a7,
23*412f47f9SXin Li };
24*412f47f9SXin Li
25*412f47f9SXin Li static svfloat32_t NOINLINE
special_case(svfloat32_t x,svfloat32_t y,svbool_t pg)26*412f47f9SXin Li special_case (svfloat32_t x, svfloat32_t y, svbool_t pg)
27*412f47f9SXin Li {
28*412f47f9SXin Li return sv_call_f32 (sinhf, x, y, pg);
29*412f47f9SXin Li }
30*412f47f9SXin Li
31*412f47f9SXin Li /* Approximation for SVE single-precision sinh(x) using expm1.
32*412f47f9SXin Li sinh(x) = (exp(x) - exp(-x)) / 2.
33*412f47f9SXin Li The maximum error is 2.26 ULP:
34*412f47f9SXin Li _ZGVsMxv_sinhf (0x1.e34a9ep-4) got 0x1.e469ep-4
35*412f47f9SXin Li want 0x1.e469e4p-4. */
SV_NAME_F1(sinh)36*412f47f9SXin Li svfloat32_t SV_NAME_F1 (sinh) (svfloat32_t x, const svbool_t pg)
37*412f47f9SXin Li {
38*412f47f9SXin Li const struct data *d = ptr_barrier (&data);
39*412f47f9SXin Li svfloat32_t ax = svabs_x (pg, x);
40*412f47f9SXin Li svuint32_t sign
41*412f47f9SXin Li = sveor_x (pg, svreinterpret_u32 (x), svreinterpret_u32 (ax));
42*412f47f9SXin Li svfloat32_t halfsign = svreinterpret_f32 (svorr_x (pg, sign, d->halff));
43*412f47f9SXin Li
44*412f47f9SXin Li svbool_t special = svcmpge (pg, svreinterpret_u32 (ax), d->large_bound);
45*412f47f9SXin Li
46*412f47f9SXin Li /* Up to the point that expm1f overflows, we can use it to calculate sinhf
47*412f47f9SXin Li using a slight rearrangement of the definition of asinh. This allows us to
48*412f47f9SXin Li retain acceptable accuracy for very small inputs. */
49*412f47f9SXin Li svfloat32_t t = expm1f_inline (ax, pg, &d->expm1f_consts);
50*412f47f9SXin Li t = svadd_x (pg, t, svdiv_x (pg, t, svadd_x (pg, t, 1.0)));
51*412f47f9SXin Li
52*412f47f9SXin Li /* Fall back to the scalar variant for any lanes which would cause
53*412f47f9SXin Li expm1f to overflow. */
54*412f47f9SXin Li if (unlikely (svptest_any (pg, special)))
55*412f47f9SXin Li return special_case (x, svmul_x (pg, t, halfsign), special);
56*412f47f9SXin Li
57*412f47f9SXin Li return svmul_x (pg, t, halfsign);
58*412f47f9SXin Li }
59*412f47f9SXin Li
60*412f47f9SXin Li PL_SIG (SV, F, 1, sinh, -10.0, 10.0)
61*412f47f9SXin Li PL_TEST_ULP (SV_NAME_F1 (sinh), 1.76)
62*412f47f9SXin Li PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinh), 0, 0x1.6a09e8p-32, 1000)
63*412f47f9SXin Li PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinh), 0x1.6a09e8p-32, 0x42b0c0a7, 100000)
64*412f47f9SXin Li PL_TEST_SYM_INTERVAL (SV_NAME_F1 (sinh), 0x42b0c0a7, inf, 1000)
65