1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
|
/*
* Single-precision vector sinh(x) function.
* Copyright (c) 2022, Arm Limited.
* SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
*/
#include "v_math.h"
#include "mathlib.h"
#include "pl_sig.h"
#include "pl_test.h"
#if V_SUPPORTED
#define AbsMask 0x7fffffff
#define Half 0x3f000000
#define Expm1OFlowLimit \
0x42b17218 /* 0x1.62e43p+6, 2^7*ln2, minimum value for which expm1f \
overflows. */
/* Approximation for vector single-precision sinh(x) using expm1.
sinh(x) = (exp(x) - exp(-x)) / 2.
The maximum error is 2.26 ULP:
__v_sinhf(0x1.e34a9ep-4) got 0x1.e469ep-4 want 0x1.e469e4p-4. */
VPCS_ATTR v_f32_t V_NAME (sinhf) (v_f32_t x)
{
v_u32_t ix = v_as_u32_f32 (x);
v_u32_t iax = ix & AbsMask;
v_f32_t ax = v_as_f32_u32 (iax);
v_u32_t sign = ix & ~AbsMask;
v_f32_t halfsign = v_as_f32_u32 (sign | Half);
v_u32_t special = v_cond_u32 (iax >= Expm1OFlowLimit);
/* Fall back to the scalar variant for all lanes if any of them should trigger
an exception. */
if (unlikely (v_any_u32 (special)))
return v_call_f32 (sinhf, x, x, v_u32 (-1));
/* Up to the point that expm1f overflows, we can use it to calculate sinhf
using a slight rearrangement of the definition of asinh. This allows us to
retain acceptable accuracy for very small inputs. */
v_f32_t t = V_NAME (expm1f) (ax);
return (t + t / (t + 1)) * halfsign;
}
VPCS_ALIAS
PL_SIG (V, F, 1, sinh, -10.0, 10.0)
PL_TEST_ULP (V_NAME (sinhf), 1.76)
PL_TEST_EXPECT_FENV (V_NAME (sinhf), WANT_ERRNO)
PL_TEST_INTERVAL (V_NAME (sinhf), 0, 0x1.62e43p+6, 100000)
PL_TEST_INTERVAL (V_NAME (sinhf), -0, -0x1.62e43p+6, 100000)
PL_TEST_INTERVAL (V_NAME (sinhf), 0x1.62e43p+6, 0x1.65a9fap+6, 100)
PL_TEST_INTERVAL (V_NAME (sinhf), -0x1.62e43p+6, -0x1.65a9fap+6, 100)
PL_TEST_INTERVAL (V_NAME (sinhf), 0x1.65a9fap+6, inf, 100)
PL_TEST_INTERVAL (V_NAME (sinhf), -0x1.65a9fap+6, -inf, 100)
#endif
|
