Am 30.09.2022 um 16:31 schrieb Riccardo (Jack) Lucchetti:
On Fri, 30 Sep 2022, Sven Schreiber wrote:
> Jack, maybe you could give a concrete example (with our StrucTiSM,
> perhaps?) of what kind of gains we're talking about.
It really depends on the kind of problem you have and on your hw/sw
combination. For example, I'm attaching a zip file with 3 scripts: one
uses StrucTiSM to estimate a BSM with some exogenous variables, one is
a slight modification of the "Uhlig" example in the SVAR package and
the third one is a Jesús Fernández-Villaverde-style RBC script. On my
work PC (linux with 6 Intel i7-8700 cores) I'm getting the following
bikes Uhlig RBC
2022b 1.73981 11.0784 189.335
2022c-git 1.61362 9.97556 56.1038
speed gain 7.8% 11.1% 237.5%
So you can see how different the time savings can be. Numerical
results are, of course, identical.
[I'm switching this to the devel list.]
Yes, for your RBC script ("...minifunc...") I'm seeing a change of
runtimes from 720 secs to about 140 seconds on a Win10-laptop here, so
that's certainly very nice.
However, consider a similar exercise in my protoype package (from 2020)
which I'm attaching. That "hanslspeedjulia" package comes with an
"examples" directory which has a RBC_Hansldriver.inp script. If you open
that script and change the line "tryJulia = 1" to "tryJulia = 0", then
it runs a hansl function which is also based on Federico's earlier RBC
code. But with that thing I'm _not_ seeing a difference between gretl
2022b and the current snapshot. Obviously the scripts are not identical,
but I don't have time right now to analyze where the difference is.
Off-topic: This package of mine hasn't been published because it relies
on an external program (Julia, of course). Some time ago we discussed
that perhaps it might still be published because it provides the
translation of (a subset of) hansl to Julia code. And the actual
invocation of Julia would not be happening in the sample script and so
"officially" a Julia installation would not be needed; instead the
invocation of Julia would only be documented and described, and examples
would be given. This never happened, however, partly because it feels a
little bit like cheating. So the question right now is, should such a
"trick publishing" of this package happen, anyway?