1*f3087befSAndrew Turner /*
2*f3087befSAndrew Turner * Single-precision vector sincos function.
3*f3087befSAndrew Turner *
4*f3087befSAndrew Turner * Copyright (c) 2023-2024, Arm Limited.
5*f3087befSAndrew Turner * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception
6*f3087befSAndrew Turner */
7*f3087befSAndrew Turner
8*f3087befSAndrew Turner /* Define _GNU_SOURCE in order to include sincosf declaration. If building
9*f3087befSAndrew Turner pre-GLIBC 2.1, or on a non-GNU conforming system, this routine will need to
10*f3087befSAndrew Turner be linked against the scalar sincosf from math/. */
11*f3087befSAndrew Turner #define _GNU_SOURCE
12*f3087befSAndrew Turner #include <math.h>
13*f3087befSAndrew Turner
14*f3087befSAndrew Turner #include "v_sincosf_common.h"
15*f3087befSAndrew Turner #include "v_math.h"
16*f3087befSAndrew Turner #include "test_defs.h"
17*f3087befSAndrew Turner
18*f3087befSAndrew Turner /* sincos not available for all scalar libm implementations. */
19*f3087befSAndrew Turner #if defined(_MSC_VER) || !defined(__GLIBC__)
20*f3087befSAndrew Turner static void
sincosf(float x,float * out_sin,float * out_cos)21*f3087befSAndrew Turner sincosf (float x, float *out_sin, float *out_cos)
22*f3087befSAndrew Turner {
23*f3087befSAndrew Turner *out_sin = sinf (x);
24*f3087befSAndrew Turner *out_cos = cosf (x);
25*f3087befSAndrew Turner }
26*f3087befSAndrew Turner #endif
27*f3087befSAndrew Turner
28*f3087befSAndrew Turner static void VPCS_ATTR NOINLINE
special_case(float32x4_t x,uint32x4_t special,float * out_sin,float * out_cos)29*f3087befSAndrew Turner special_case (float32x4_t x, uint32x4_t special, float *out_sin,
30*f3087befSAndrew Turner float *out_cos)
31*f3087befSAndrew Turner {
32*f3087befSAndrew Turner for (int i = 0; i < 4; i++)
33*f3087befSAndrew Turner if (special[i])
34*f3087befSAndrew Turner sincosf (x[i], out_sin + i, out_cos + i);
35*f3087befSAndrew Turner }
36*f3087befSAndrew Turner
37*f3087befSAndrew Turner /* Single-precision vector function allowing calculation of both sin and cos in
38*f3087befSAndrew Turner one function call, using shared argument reduction and separate low-order
39*f3087befSAndrew Turner polynomials.
40*f3087befSAndrew Turner Worst-case error for sin is 1.67 ULP:
41*f3087befSAndrew Turner v_sincosf_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5
42*f3087befSAndrew Turner Worst-case error for cos is 1.81 ULP:
43*f3087befSAndrew Turner v_sincosf_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6. */
44*f3087befSAndrew Turner VPCS_ATTR void
_ZGVnN4vl4l4_sincosf(float32x4_t x,float * out_sin,float * out_cos)45*f3087befSAndrew Turner _ZGVnN4vl4l4_sincosf (float32x4_t x, float *out_sin, float *out_cos)
46*f3087befSAndrew Turner {
47*f3087befSAndrew Turner const struct v_sincosf_data *d = ptr_barrier (&v_sincosf_data);
48*f3087befSAndrew Turner uint32x4_t special = check_ge_rangeval (x, d);
49*f3087befSAndrew Turner
50*f3087befSAndrew Turner float32x4x2_t sc = v_sincosf_inline (x, d);
51*f3087befSAndrew Turner
52*f3087befSAndrew Turner vst1q_f32 (out_sin, sc.val[0]);
53*f3087befSAndrew Turner vst1q_f32 (out_cos, sc.val[1]);
54*f3087befSAndrew Turner
55*f3087befSAndrew Turner if (unlikely (v_any_u32 (special)))
56*f3087befSAndrew Turner special_case (x, special, out_sin, out_cos);
57*f3087befSAndrew Turner }
58*f3087befSAndrew Turner
59*f3087befSAndrew Turner TEST_DISABLE_FENV (_ZGVnN4v_sincosf_sin)
60*f3087befSAndrew Turner TEST_DISABLE_FENV (_ZGVnN4v_sincosf_cos)
61*f3087befSAndrew Turner TEST_ULP (_ZGVnN4v_sincosf_sin, 1.17)
62*f3087befSAndrew Turner TEST_ULP (_ZGVnN4v_sincosf_cos, 1.31)
63*f3087befSAndrew Turner #define V_SINCOSF_INTERVAL(lo, hi, n) \
64*f3087befSAndrew Turner TEST_INTERVAL (_ZGVnN4v_sincosf_sin, lo, hi, n) \
65*f3087befSAndrew Turner TEST_INTERVAL (_ZGVnN4v_sincosf_cos, lo, hi, n)
66*f3087befSAndrew Turner V_SINCOSF_INTERVAL (0, 0x1p-31, 50000)
67*f3087befSAndrew Turner V_SINCOSF_INTERVAL (0x1p-31, 0x1p20, 500000)
68*f3087befSAndrew Turner V_SINCOSF_INTERVAL (0x1p20, inf, 10000)
69