xref: /freebsd/contrib/arm-optimized-routines/pl/math/v_cexpif_1u8.c (revision 7fdf597e96a02165cfe22ff357b857d5fa15ed8a)
1 /*
2  * Single-precision vector cexpi function.
3  *
4  * Copyright (c) 2023, Arm Limited.
5  * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6  */
7 
8 #include "v_sincosf_common.h"
9 #include "v_math.h"
10 #include "pl_test.h"
11 
12 static float32x4x2_t VPCS_ATTR NOINLINE
13 special_case (float32x4_t x, uint32x4_t special, float32x4x2_t y)
14 {
15   return (float32x4x2_t){ v_call_f32 (sinf, x, y.val[0], special),
16 			  v_call_f32 (cosf, x, y.val[1], special) };
17 }
18 
19 /* Single-precision vector function allowing calculation of both sin and cos in
20    one function call, using shared argument reduction and separate low-order
21    polynomials.
22    Worst-case error for sin is 1.67 ULP:
23    v_cexpif_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5
24    Worst-case error for cos is 1.81 ULP:
25    v_cexpif_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6.  */
26 VPCS_ATTR float32x4x2_t
27 _ZGVnN4v_cexpif (float32x4_t x)
28 {
29   const struct v_sincosf_data *d = ptr_barrier (&v_sincosf_data);
30   uint32x4_t special = check_ge_rangeval (x, d);
31 
32   float32x4x2_t sc = v_sincosf_inline (x, d);
33 
34   if (unlikely (v_any_u32 (special)))
35     return special_case (x, special, sc);
36   return sc;
37 }
38 
39 PL_TEST_ULP (_ZGVnN4v_cexpif_sin, 1.17)
40 PL_TEST_ULP (_ZGVnN4v_cexpif_cos, 1.31)
41 #define V_CEXPIF_INTERVAL(lo, hi, n)                                          \
42   PL_TEST_INTERVAL (_ZGVnN4v_cexpif_sin, lo, hi, n)                           \
43   PL_TEST_INTERVAL (_ZGVnN4v_cexpif_cos, lo, hi, n)
44 V_CEXPIF_INTERVAL (0, 0x1p20, 500000)
45 V_CEXPIF_INTERVAL (-0, -0x1p20, 500000)
46 V_CEXPIF_INTERVAL (0x1p20, inf, 10000)
47 V_CEXPIF_INTERVAL (-0x1p20, -inf, 10000)
48