1*412f47f9SXin Li /*
2*412f47f9SXin Li * Single-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 sincosf 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 "sv_sincosf_common.h"
16*412f47f9SXin Li #include "sv_math.h"
17*412f47f9SXin Li #include "pl_test.h"
18*412f47f9SXin Li
19*412f47f9SXin Li static void NOINLINE
special_case(svfloat32_t x,svbool_t special,float * out_sin,float * out_cos)20*412f47f9SXin Li special_case (svfloat32_t x, svbool_t special, float *out_sin, float *out_cos)
21*412f47f9SXin Li {
22*412f47f9SXin Li svbool_t p = svptrue_pat_b32 (SV_VL1);
23*412f47f9SXin Li for (int i = 0; i < svcntw (); i++)
24*412f47f9SXin Li {
25*412f47f9SXin Li if (svptest_any (special, p))
26*412f47f9SXin Li sincosf (svlastb (p, x), out_sin + i, out_cos + i);
27*412f47f9SXin Li p = svpnext_b32 (svptrue_b32 (), p);
28*412f47f9SXin Li }
29*412f47f9SXin Li }
30*412f47f9SXin Li
31*412f47f9SXin Li /* Single-precision vector function allowing calculation of both sin and cos in
32*412f47f9SXin Li one function call, using shared argument reduction and separate low-order
33*412f47f9SXin Li polynomials.
34*412f47f9SXin Li Worst-case error for sin is 1.67 ULP:
35*412f47f9SXin Li sv_sincosf_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5
36*412f47f9SXin Li Worst-case error for cos is 1.81 ULP:
37*412f47f9SXin Li sv_sincosf_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6. */
38*412f47f9SXin Li void
_ZGVsMxvl4l4_sincosf(svfloat32_t x,float * out_sin,float * out_cos,svbool_t pg)39*412f47f9SXin Li _ZGVsMxvl4l4_sincosf (svfloat32_t x, float *out_sin, float *out_cos,
40*412f47f9SXin Li svbool_t pg)
41*412f47f9SXin Li {
42*412f47f9SXin Li const struct sv_sincosf_data *d = ptr_barrier (&sv_sincosf_data);
43*412f47f9SXin Li svbool_t special = check_ge_rangeval (pg, x, d);
44*412f47f9SXin Li
45*412f47f9SXin Li svfloat32x2_t sc = sv_sincosf_inline (pg, x, d);
46*412f47f9SXin Li
47*412f47f9SXin Li svst1_f32 (pg, out_sin, svget2 (sc, 0));
48*412f47f9SXin Li svst1_f32 (pg, out_cos, svget2 (sc, 1));
49*412f47f9SXin Li
50*412f47f9SXin Li if (unlikely (svptest_any (pg, special)))
51*412f47f9SXin Li special_case (x, special, out_sin, out_cos);
52*412f47f9SXin Li }
53*412f47f9SXin Li
54*412f47f9SXin Li PL_TEST_ULP (_ZGVsMxv_sincosf_sin, 1.17)
55*412f47f9SXin Li PL_TEST_ULP (_ZGVsMxv_sincosf_cos, 1.31)
56*412f47f9SXin Li #define SV_SINCOSF_INTERVAL(lo, hi, n) \
57*412f47f9SXin Li PL_TEST_INTERVAL (_ZGVsMxv_sincosf_sin, lo, hi, n) \
58*412f47f9SXin Li PL_TEST_INTERVAL (_ZGVsMxv_sincosf_cos, lo, hi, n)
59*412f47f9SXin Li SV_SINCOSF_INTERVAL (0, 0x1p20, 500000)
60*412f47f9SXin Li SV_SINCOSF_INTERVAL (-0, -0x1p20, 500000)
61*412f47f9SXin Li SV_SINCOSF_INTERVAL (0x1p20, inf, 10000)
62*412f47f9SXin Li SV_SINCOSF_INTERVAL (-0x1p20, -inf, 10000)
63