SIMD Library for Evaluating Elementary Functions, vectorized libm and DFT
Boost Software License 1.0
661
stars
131
forks
source link
FP32 `log1pf_u10` and FP64 `log1p_u10` are infinitely inaccurate for large inputs #473
Open
ebavier opened 1 year ago
On x86 targets and possibly others, the
log1pf_u10
andlog1p_u10
functions are infinitely inaccurate for large input values. Currentlylog1pf_u10 (x > 1e38) = inf
, andlog1p_u10 (x > 1e307) = inf
No input value except
x = inf
should result in aninf
result forlog1p
.Relevant source:
https://github.com/shibatch/sleef/blob/85440a5e87dae36ca1b891de14bc83b441ae7c43/src/libm/sleefsp.c#L1761
https://github.com/shibatch/sleef/blob/85440a5e87dae36ca1b891de14bc83b441ae7c43/src/libm/sleefdp.c#L2179