This is the mail archive of the fortran@gcc.gnu.org mailing list for the GNU Fortran project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: Polyhedron Benchmark


Hello!

> Could you perhaps add -fvect-cost-model to
> gcc-4.3 compile flags? There are some surprises there ;)

Well, a 2% overall improvement, some slower, some faster. Only for
induct an improvement by a 1/3.

Should I use -fvect-cost-model in general for my nightly runs or should
I add another column with one -fvect-cost-model run?

I think that polyhedron could be a nice head-to-head comparison for builds with and without -fvect-cost-model. Since I expect that there will be quite some work needed to tune the vectorizer - and if you can spare some cycles - I'd propose to add another column to the table.


BTW: Looking at your benchmark page, you have links to some stalled PRs there. PR32457 was fixed some time ago and PR32084 (and IMO also 32086) is due to "vectorize-all-or-nothing" approach implemented by current vectorizer (should be fixed by -fvect-cost-model).

For Core2 the results with -fvect-cost-model are:

I _think_ that on K8, the difference was even larger (current X86_64 vect cost model is written for K8, not Core2), but it was some time ago that I have played with this option.


Uros.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]