1*f3087befSAndrew Turner /* 2*f3087befSAndrew Turner * Single-precision vector cexpi 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 #include "v_sincosf_common.h" 9*f3087befSAndrew Turner #include "v_math.h" 10*f3087befSAndrew Turner #include "test_defs.h" 11*f3087befSAndrew Turner 12*f3087befSAndrew Turner static float32x4x2_t VPCS_ATTR NOINLINE 13*f3087befSAndrew Turner special_case (float32x4_t x, uint32x4_t special, float32x4x2_t y) 14*f3087befSAndrew Turner { 15*f3087befSAndrew Turner return (float32x4x2_t){ v_call_f32 (sinf, x, y.val[0], special), 16*f3087befSAndrew Turner v_call_f32 (cosf, x, y.val[1], special) }; 17*f3087befSAndrew Turner } 18*f3087befSAndrew Turner 19*f3087befSAndrew Turner /* Single-precision vector function allowing calculation of both sin and cos in 20*f3087befSAndrew Turner one function call, using shared argument reduction and separate low-order 21*f3087befSAndrew Turner polynomials. 22*f3087befSAndrew Turner Worst-case error for sin is 1.67 ULP: 23*f3087befSAndrew Turner v_cexpif_sin(0x1.c704c4p+19) got 0x1.fff698p-5 want 0x1.fff69cp-5 24*f3087befSAndrew Turner Worst-case error for cos is 1.81 ULP: 25*f3087befSAndrew Turner v_cexpif_cos(0x1.e506fp+19) got -0x1.ffec6ep-6 want -0x1.ffec72p-6. */ 26*f3087befSAndrew Turner VPCS_ATTR float32x4x2_t 27*f3087befSAndrew Turner _ZGVnN4v_cexpif (float32x4_t x) 28*f3087befSAndrew Turner { 29*f3087befSAndrew Turner const struct v_sincosf_data *d = ptr_barrier (&v_sincosf_data); 30*f3087befSAndrew Turner uint32x4_t special = check_ge_rangeval (x, d); 31*f3087befSAndrew Turner 32*f3087befSAndrew Turner float32x4x2_t sc = v_sincosf_inline (x, d); 33*f3087befSAndrew Turner 34*f3087befSAndrew Turner if (unlikely (v_any_u32 (special))) 35*f3087befSAndrew Turner return special_case (x, special, sc); 36*f3087befSAndrew Turner return sc; 37*f3087befSAndrew Turner } 38*f3087befSAndrew Turner 39*f3087befSAndrew Turner TEST_DISABLE_FENV (_ZGVnN4v_cexpif_sin) 40*f3087befSAndrew Turner TEST_DISABLE_FENV (_ZGVnN4v_cexpif_cos) 41*f3087befSAndrew Turner TEST_ULP (_ZGVnN4v_cexpif_sin, 1.17) 42*f3087befSAndrew Turner TEST_ULP (_ZGVnN4v_cexpif_cos, 1.31) 43*f3087befSAndrew Turner #define V_CEXPIF_INTERVAL(lo, hi, n) \ 44*f3087befSAndrew Turner TEST_INTERVAL (_ZGVnN4v_cexpif_sin, lo, hi, n) \ 45*f3087befSAndrew Turner TEST_INTERVAL (_ZGVnN4v_cexpif_cos, lo, hi, n) 46*f3087befSAndrew Turner V_CEXPIF_INTERVAL (0, 0x1p20, 500000) 47*f3087befSAndrew Turner V_CEXPIF_INTERVAL (-0, -0x1p20, 500000) 48*f3087befSAndrew Turner V_CEXPIF_INTERVAL (0x1p20, inf, 10000) 49*f3087befSAndrew Turner V_CEXPIF_INTERVAL (-0x1p20, -inf, 10000) 50