Re: [Gretl-devel] [Gretl-users] When Julia is faster than C and Gretl

[taking this to devel instead of users] Am 13.02.2018 um 18:23 schrieb Allin Cottrell: OK, since I still don't have much experience with Julia I have played around with Python/Numba on the stationary bootstrap (SB) first. I'm attaching a horserace between the default SB function from SB.gfn and a JITted Python/Numpy/Numba implementation of the loop-based version SB_old2. It turns out Numba is 5x faster for the pure calculations (well, depending on the problem size due to the fixed overhead), but the bottleneck is the disk-based transfer of the big random matrix back to the gretl environment. Here are my results for 39999 replications: SB 1.3, 2017-11-14 (F. Di Iorio, S. Fachin, A. Tarassow, Riccardo "Jack" Lucchetti) Gretl native (SB package) This took = 6.067643 sec. Python / Numba pure calculation This took = 1.323467 sec. Python / Numba with transfer This took = 7.564688 sec. I believe something similar would emerge for Julia. So unless the data transfer can be organized in some other way I think it's only possible to beat well-written hansl code when the data reduction step can already happen on the Python (or Julia) side. cheers, sven