Path: blob/main/contrib/arm-optimized-routines/math/aarch64/advsimd/log1p.c
48378 views
/*1* Double-precision vector log(1+x) function.2*3* Copyright (c) 2022-2024, Arm Limited.4* SPDX-License-Identifier: MIT OR Apache-2.0 WITH LLVM-exception5*/67#include "v_math.h"8#include "test_sig.h"9#include "test_defs.h"1011#define WANT_V_LOG1P_K0_SHORTCUT 012#include "v_log1p_inline.h"1314const static struct data15{16struct v_log1p_data d;17uint64x2_t inf, minus_one;18} data = { .d = V_LOG1P_CONSTANTS_TABLE,19.inf = V2 (0x7ff0000000000000),20.minus_one = V2 (0xbff0000000000000) };2122#define BottomMask v_u64 (0xffffffff)2324static float64x2_t NOINLINE VPCS_ATTR25special_case (float64x2_t x, uint64x2_t cmp, const struct data *d)26{27/* Side-step special lanes so fenv exceptions are not triggered28inadvertently. */29float64x2_t x_nospecial = v_zerofy_f64 (x, cmp);30return v_call_f64 (log1p, x, log1p_inline (x_nospecial, &d->d), cmp);31}3233/* Vector log1p approximation using polynomial on reduced interval. Routine is34a modification of the algorithm used in scalar log1p, with no shortcut for35k=0 and no narrowing for f and k. Maximum observed error is 2.45 ULP:36_ZGVnN2v_log1p(0x1.658f7035c4014p+11) got 0x1.fd61d0727429dp+237want 0x1.fd61d0727429fp+2 . */38VPCS_ATTR float64x2_t V_NAME_D1 (log1p) (float64x2_t x)39{40const struct data *d = ptr_barrier (&data);41uint64x2_t ix = vreinterpretq_u64_f64 (x);42uint64x2_t ia = vreinterpretq_u64_f64 (vabsq_f64 (x));4344uint64x2_t special_cases45= vorrq_u64 (vcgeq_u64 (ia, d->inf), vcgeq_u64 (ix, d->minus_one));4647if (unlikely (v_any_u64 (special_cases)))48return special_case (x, special_cases, d);4950return log1p_inline (x, &d->d);51}5253TEST_SIG (V, D, 1, log1p, -0.9, 10.0)54TEST_ULP (V_NAME_D1 (log1p), 1.95)55TEST_DISABLE_FENV_IF_NOT (V_NAME_D1 (log1p), WANT_SIMD_EXCEPT)56TEST_SYM_INTERVAL (V_NAME_D1 (log1p), 0.0, 0x1p-23, 50000)57TEST_SYM_INTERVAL (V_NAME_D1 (log1p), 0x1p-23, 0.001, 50000)58TEST_SYM_INTERVAL (V_NAME_D1 (log1p), 0.001, 1.0, 50000)59TEST_INTERVAL (V_NAME_D1 (log1p), 1, inf, 40000)60TEST_INTERVAL (V_NAME_D1 (log1p), -1.0, -inf, 500)616263