1 /* 2 * Single-precision vector sincos function. 3 * 4 * Copyright (c) 2023, Arm Limited. 5 * SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception 6 */ 7 8 /* Define _GNU_SOURCE in order to include sincosf declaration. If building 9 pre-GLIBC 2.1, or on a non-GNU conforming system, this routine will need to 10 be linked against the scalar sincosf from math/. */ 11 #define _GNU_SOURCE 12 #include <math.h> 13 #undef _GNU_SOURCE 14 15 #include "v_sincosf_common.h" 16 #include "v_math.h" 17 #include "pl_test.h" 18 19 static void VPCS_ATTR NOINLINE 20 special_case (float32x4_t x, uint32x4_t special, float *out_sin, 21 float *out_cos) 22 { 23 for (int i = 0; i < 4; i++) 24 if (special[i]) 25 sincosf (x[i], out_sin + i, out_cos + i); 26 } 27 28 /* Single-precision vector function allowing calculation of both sin and cos in 29 one function call, using shared argument reduction and separate low-order 30 polynomials. 31 Worst-case error for sin is 1.67 ULP: 32 v_sincosf_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5 33 Worst-case error for cos is 1.81 ULP: 34 v_sincosf_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6. */ 35 VPCS_ATTR void 36 _ZGVnN4vl4l4_sincosf (float32x4_t x, float *out_sin, float *out_cos) 37 { 38 const struct v_sincosf_data *d = ptr_barrier (&v_sincosf_data); 39 uint32x4_t special = check_ge_rangeval (x, d); 40 41 float32x4x2_t sc = v_sincosf_inline (x, d); 42 43 vst1q_f32 (out_sin, sc.val[0]); 44 vst1q_f32 (out_cos, sc.val[1]); 45 46 if (unlikely (v_any_u32 (special))) 47 special_case (x, special, out_sin, out_cos); 48 } 49 50 PL_TEST_ULP (_ZGVnN4v_sincosf_sin, 1.17) 51 PL_TEST_ULP (_ZGVnN4v_sincosf_cos, 1.31) 52 #define V_SINCOSF_INTERVAL(lo, hi, n) \ 53 PL_TEST_INTERVAL (_ZGVnN4v_sincosf_sin, lo, hi, n) \ 54 PL_TEST_INTERVAL (_ZGVnN4v_sincosf_cos, lo, hi, n) 55 V_SINCOSF_INTERVAL (0, 0x1p20, 500000) 56 V_SINCOSF_INTERVAL (-0, -0x1p20, 500000) 57 V_SINCOSF_INTERVAL (0x1p20, inf, 10000) 58 V_SINCOSF_INTERVAL (-0x1p20, -inf, 10000) 59