[Bug libfortran/51119] MATMUL slow for large matrices

[Joost.VandeVondele at mat dot ethz.ch]

http://gcc.gnu.org/bugzilla/show_bug.cgi?id=51119

Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |Joost.VandeVondele at mat
                   |                            |dot ethz.ch

--- Comment #8 from Joost VandeVondele <Joost.VandeVondele at mat dot ethz.ch> 2012-06-29 07:19:03 UTC ---
(In reply to comment #7)
> (In reply to comment #6)
> > Janne, have you had a chance to look at this ? For larger matrices MATMMUL is
> > really slow. Anything that includes even the most basic blocking scheme should
> > be faster. I think this would be a valuable improvement.
> 
> I implemented a block-panel multiplication algorithm similar to GOTO BLAS and
> Eigen, but I got side-tracked by other things and never found the time to fix
> the corner-case bugs and tune performance. IIRC I reached about 30-40 % of peak
> flops which was a bit disappointing.

I think 30% of peak is a good improvement over the current version (which
reaches 7% of peak (92% for MKL) for a double precision 8000x8000 matrix
multiplication) on a sandy bridge.

In addition to blocking, is the Fortran runtime being compiled with a set of
compile options that enables vectorization ? In the ideal world, gcc would
recognize the loop pattern in the runtime library code, and do blocking,
vectorization etc. automagically.