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