[Nauty] Squeezing out another 10%...
Brendan McKay
bdm at cs.anu.edu.au
Mon Feb 14 18:30:02 EST 2005
* Gordon Royle <gordon at csse.uwa.edu.au> [050214 17:04]:
>
> I have discovered through experimentation that I get an additional 10%
> performance out of geng (and hence presumably) nauty by altering the
> optimization flags from the supplied defaults... I am using a Pentium 4
> machine, and if I put
>
> gcc -O2 -Dmarch=pentium4
I guess you mean
gcc -O2 -march=pentium4
Somehow the configuration file (configure.ac) misses this case.
Now it is included.
I use an athlon k7, but gcc 3.3.4 does not have -march=k7. Using
-march=k6 gives worse performance than the default -march=i686.
Trying to automatically select the best options is hard.
> then it is about 10% faster than
>
> gcc -O4
>
> Interestingly, it proves better to use O1/O2 with the
> architecture-specific flag than to use O3 or higher...
That will probably depend on which exact chip you have (eg specs
of on-board cache) and will also vary from one program to another.
Brendan.
More information about the Nauty
mailing list