Polyhedron 2005 Benchmarks

[Daniel Berlin]

On Fri, 2005-11-18 at 15:19 +0100, Dominique Dhumieres wrote:
> > Shouldn't -ftree-loop-linear perform the loop interchange here?
> 
> [karma] lin/source% gfortran -O3 -ffast-math -funroll-loops air.f90
> [karma] lin/source% time a.out > tmp
> 44.720u 0.580s 0:47.92 94.5%    0+0k 0+12io 0pf+0w
> [karma] lin/source% gfortran -O3 -ftree-loop-linear -ffast-math -funroll-loops air.f90
> [karma] lin/source% time a.out > tmp
> 31.010u 0.590s 0:35.14 89.9%    0+0k 0+10io 0pf+0w
> 
> At least a visible effect. Now the surprising part is with a variant in which I have
> done half a dozen permutations manually:
> 
> [karma] lin/source% gfortran -O3 -ffast-math -funroll-loops air.v1.f90
> [karma] lin/source% time a.out > tmp
> 24.690u 0.370s 0:26.93 93.0%    0+0k 0+9io 0pf+0w
> [karma] lin/source% gfortran -O3 -ftree-loop-linear -ffast-math -funroll-loops air.v1.f90
> [karma] lin/source% time a.out > tmp
> 17.380u 0.350s 0:18.95 93.5%    0+0k 0+12io 0pf+0w
> 
> > I'd be curious to know if it decides not to, and why.
> > -fdump-tree-ltrans-all
> 
> added to gfortran -O3 -ftree-loop-linear -ffast-math -funroll-loops air.v1.f90
> it produces a 6369030 long file air.v1.f90.t75.ltrans. If someone is interested
> I can do it for the othercases and put the results in my web pages.
> 

If you could gzip/bzip2 that file and send it to me, i'd still
appreciate it
> Dominique