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