1*412f47f9SXin Li /*
2*412f47f9SXin Li * Single-precision vector cexpi 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 "v_sincosf_common.h"
9*412f47f9SXin Li #include "v_math.h"
10*412f47f9SXin Li #include "pl_test.h"
11*412f47f9SXin Li
12*412f47f9SXin Li static float32x4x2_t VPCS_ATTR NOINLINE
special_case(float32x4_t x,uint32x4_t special,float32x4x2_t y)13*412f47f9SXin Li special_case (float32x4_t x, uint32x4_t special, float32x4x2_t y)
14*412f47f9SXin Li {
15*412f47f9SXin Li return (float32x4x2_t){ v_call_f32 (sinf, x, y.val[0], special),
16*412f47f9SXin Li v_call_f32 (cosf, x, y.val[1], special) };
17*412f47f9SXin Li }
18*412f47f9SXin Li
19*412f47f9SXin Li /* Single-precision vector function allowing calculation of both sin and cos in
20*412f47f9SXin Li one function call, using shared argument reduction and separate low-order
21*412f47f9SXin Li polynomials.
22*412f47f9SXin Li Worst-case error for sin is 1.67 ULP:
23*412f47f9SXin Li v_cexpif_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5
24*412f47f9SXin Li Worst-case error for cos is 1.81 ULP:
25*412f47f9SXin Li v_cexpif_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6. */
26*412f47f9SXin Li VPCS_ATTR float32x4x2_t
_ZGVnN4v_cexpif(float32x4_t x)27*412f47f9SXin Li _ZGVnN4v_cexpif (float32x4_t x)
28*412f47f9SXin Li {
29*412f47f9SXin Li const struct v_sincosf_data *d = ptr_barrier (&v_sincosf_data);
30*412f47f9SXin Li uint32x4_t special = check_ge_rangeval (x, d);
31*412f47f9SXin Li
32*412f47f9SXin Li float32x4x2_t sc = v_sincosf_inline (x, d);
33*412f47f9SXin Li
34*412f47f9SXin Li if (unlikely (v_any_u32 (special)))
35*412f47f9SXin Li return special_case (x, special, sc);
36*412f47f9SXin Li return sc;
37*412f47f9SXin Li }
38*412f47f9SXin Li
39*412f47f9SXin Li PL_TEST_ULP (_ZGVnN4v_cexpif_sin, 1.17)
40*412f47f9SXin Li PL_TEST_ULP (_ZGVnN4v_cexpif_cos, 1.31)
41*412f47f9SXin Li #define V_CEXPIF_INTERVAL(lo, hi, n) \
42*412f47f9SXin Li PL_TEST_INTERVAL (_ZGVnN4v_cexpif_sin, lo, hi, n) \
43*412f47f9SXin Li PL_TEST_INTERVAL (_ZGVnN4v_cexpif_cos, lo, hi, n)
44*412f47f9SXin Li V_CEXPIF_INTERVAL (0, 0x1p20, 500000)
45*412f47f9SXin Li V_CEXPIF_INTERVAL (-0, -0x1p20, 500000)
46*412f47f9SXin Li V_CEXPIF_INTERVAL (0x1p20, inf, 10000)
47*412f47f9SXin Li V_CEXPIF_INTERVAL (-0x1p20, -inf, 10000)
48