Re: [gnugo-devel] double dfas and prerotated dfas

[Paul Pogonyshev]

Evan wrote:
> On Sun, 3 Aug 2003, Paul Pogonyshev wrote:
> > i have implemented "double dfa" (a dfa with two transformations
> > of each patterns) only to find that there is nothing good in it.
> > it gave less than 1% speedup at the cost of 2.5 - 3 times dfa
> > size increase.  with computers becoming faster this will surely
> > be a disadvantage and eventually double dfa will be slower.
>
> What was the relative speedup of the DFA code? how big a piece of total
> runtime is the DFA code?  If the DFA code becomes a larger part of total
> runtime, it might matter; my understanding is that the DFA is a larger
> part of total runtime than it used to be, but I haven't verified this.

i didn't measure.  double dfa was indeed _faster_, but the difference was
so small that i decided it wasn't worth the cost.

in latest profiles i have (3.3.22 maybe, don't remember), dfa code uses
7 - 11 % of total runtime depending on test suite.

Paul