Re: [gnugo-devel] engine/influence.c (and DFA)

[Dave Denholm]

Gunnar Farneback <address@hidden> writes:

> Dave wrote:
> >    I was just having a look at influence.c, since that is where the
> > profile shows the most time being spent.
> 
> On what platform and what exactly did you run? I just did a profile on
> an AMD Athlon running neurogo.tst and then the influence code doesn't
> take much time. There the pattern matching is in the top.


Hmm - well, it was the latest tar file available (3.3.6 I think) ,
on a pentium III (portable), probably running   ./gnugo -b 500 
or something like that.

I'll run it again at home (I'm at work right now) and mail the result
I got.


*However*, I realised I should have mentioned that the numbers I was posting
today were based on a very old version (3.1.3), since that's what I had
on the machine at work, and there has since been a policy change such
that I can no longer download new code onto a machine at work.

I checked that the influence code was still similar to the latest, but
the speedups are from the old code. Sorry, I should have made that clear.


> 
> What we see here is probably that the neurogo test suite is rather owl
> intensive. For a more representative profiling one should probably
> replay a couple of entire games against varied opponents (self play
> just doesn't suffice). The most straightforward way to do that would
> be to have a GTP file simulating replay of a selection of games.
> 


Hmm - good idea - perhaps for raw micro-optimisation work, there should
be a such a standard timing script. (Unless there already is one - as
usual, I'm very out of touch with latest developments)


I assume there are already scripts to ask gnugo to generate a move at
each position of a game from an sgf file ?


dd
-- 
address@hidden          http://www.insignia.com