1*412f47f9SXin Li /*
2*412f47f9SXin Li * Double-precision vector sincos 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 /* Define _GNU_SOURCE in order to include sincos declaration. If building
9*412f47f9SXin Li pre-GLIBC 2.1, or on a non-GNU conforming system, this routine will need to
10*412f47f9SXin Li be linked against the scalar sincosf from math/. */
11*412f47f9SXin Li #define _GNU_SOURCE
12*412f47f9SXin Li #include <math.h>
13*412f47f9SXin Li #undef _GNU_SOURCE
14*412f47f9SXin Li
15*412f47f9SXin Li #include "v_math.h"
16*412f47f9SXin Li #include "pl_test.h"
17*412f47f9SXin Li #include "v_sincos_common.h"
18*412f47f9SXin Li
19*412f47f9SXin Li static void VPCS_ATTR NOINLINE
special_case(float64x2_t x,uint64x2_t special,double * out_sin,double * out_cos)20*412f47f9SXin Li special_case (float64x2_t x, uint64x2_t special, double *out_sin,
21*412f47f9SXin Li double *out_cos)
22*412f47f9SXin Li {
23*412f47f9SXin Li if (special[0])
24*412f47f9SXin Li sincos (x[0], out_sin, out_cos);
25*412f47f9SXin Li if (special[1])
26*412f47f9SXin Li sincos (x[1], out_sin + 1, out_cos + 1);
27*412f47f9SXin Li }
28*412f47f9SXin Li
29*412f47f9SXin Li /* Double-precision vector function allowing calculation of both sin and cos in
30*412f47f9SXin Li one function call, using shared argument reduction and separate polynomials.
31*412f47f9SXin Li Largest observed error is for sin, 3.22 ULP:
32*412f47f9SXin Li v_sincos_sin (0x1.d70eef40f39b1p+12) got -0x1.ffe9537d5dbb7p-3
33*412f47f9SXin Li want -0x1.ffe9537d5dbb4p-3. */
34*412f47f9SXin Li VPCS_ATTR void
_ZGVnN2vl8l8_sincos(float64x2_t x,double * out_sin,double * out_cos)35*412f47f9SXin Li _ZGVnN2vl8l8_sincos (float64x2_t x, double *out_sin, double *out_cos)
36*412f47f9SXin Li {
37*412f47f9SXin Li const struct v_sincos_data *d = ptr_barrier (&v_sincos_data);
38*412f47f9SXin Li uint64x2_t special = check_ge_rangeval (x, d);
39*412f47f9SXin Li
40*412f47f9SXin Li float64x2x2_t sc = v_sincos_inline (x, d);
41*412f47f9SXin Li
42*412f47f9SXin Li vst1q_f64 (out_sin, sc.val[0]);
43*412f47f9SXin Li vst1q_f64 (out_cos, sc.val[1]);
44*412f47f9SXin Li
45*412f47f9SXin Li if (unlikely (v_any_u64 (special)))
46*412f47f9SXin Li special_case (x, special, out_sin, out_cos);
47*412f47f9SXin Li }
48*412f47f9SXin Li
49*412f47f9SXin Li PL_TEST_ULP (_ZGVnN2v_sincos_sin, 2.73)
50*412f47f9SXin Li PL_TEST_ULP (_ZGVnN2v_sincos_cos, 2.73)
51*412f47f9SXin Li #define V_SINCOS_INTERVAL(lo, hi, n) \
52*412f47f9SXin Li PL_TEST_INTERVAL (_ZGVnN2v_sincos_sin, lo, hi, n) \
53*412f47f9SXin Li PL_TEST_INTERVAL (_ZGVnN2v_sincos_cos, lo, hi, n)
54*412f47f9SXin Li V_SINCOS_INTERVAL (0, 0x1p23, 500000)
55*412f47f9SXin Li V_SINCOS_INTERVAL (-0, -0x1p23, 500000)
56*412f47f9SXin Li V_SINCOS_INTERVAL (0x1p23, inf, 10000)
57*412f47f9SXin Li V_SINCOS_INTERVAL (-0x1p23, -inf, 10000)
58