Re: [Gretl-devel] simulation speed
On Wed, 23 Dec 2009, Sven Schreiber wrote:
Talha Yalta schrieb:
>> ...I have "borrowed" the code Jochen
>> Voss wrote for gsl to implement the ziggurat method for generating
>> random normal values. Here's what I'm seeing after making this
>> change:
>>
>> loop vectorized
>> octave 2.59 0.53
>> gretl 0.87 0.39
>> ox 0.85 0.56
> Wow!
>
Yes I was gonna say that as well!
There is a possible catch: Doornik has an article (google: doornik
ziggurat) arguing that the standard ziggurat method fails a test
for high-dimensional randomness (and that one might expect it to
fail), and he has a workaround which necessarily sacrifices some
of the efficiency of ziggurat. (Unfortunately his code is free
only for non-commercial use so it's incompatible with the GPL and
we can't use it in gretl.)
So far as I can tell Doornik's article hasn't been published and I
can't find any response from Marsaglia, but I'm taking this
seriously. Right now I'm running the L'Ecuyer "crush" test on the
new gretl code -- should be finished within an hour or two!
-- With respect to gsl I'm wondering what's the best
approach:
borrowing the code or requiring gsl to build gretl. I believe
you (Allin) said at some point you didn't want to have the gsl
dependency...
It's possible we may want to link to libgsl at some point, but
this particular instance wouldn't warrant that: Voss's code is
just a few dozen lines, with no integral dependence on the rest of
gsl. And the Voss code is under the GPL so borrowing is not a
problem.
Allin.