> Øystein Johansen (OJOHANS) wrote:
> >> Perhaps we should just disable the mt speed test, until
> >> somebody fixes it.
> > Yes! Remove it until further. The speed test should be reworked
> completely. Today it generates a random position and then evaluates
> the position -- and repeats this. However, picking a random position
> is quite strange since you seldom get anything else than contact
> positions (Take a look at the generated positions, I dare you!). The
> speed test should have a specified set of positions which represents
> a typical match.
> > -Øystein
> Yes, we ought to have an analysed match, a few money games, and a
> rolled out positions for people to test speed and correctness on.
Having something close to a non-regression test would
be just great. First idea would be to have it embedded into gnubg's
code, maybe as a python script.
The current speed test (withot the bug) is fine enough
to measure the eval speed (which is not the overall speed you
can expect in a rollout). For example, it gives you quickly an idea of how much
a -O0 build is slower than a -03 one. I'm more for fixing it and work on
a separate non-regression part.