Re: Standard Libraries for gfortran

["Steven G. Johnson" <stevenj at alum dot mit dot edu>]

Steve Kargl wrote:
> Are you sure that optimize BLAS libraries are available 
> on *all* platforms that GCC supports?  Are these optimize
> BLAS libraries free?

ATLAS is free and supports every system that people care about for 
high-performance numerics.  For those few obscure platforms where for 
some reason it might not work (embedded systems?), people can use the 
reference implementation included with gcc.

Conversely, if you want to spend your time writing a fast matrix 
multiply for any given platform, you should write something that 
conforms to the BLAS API, not some internal API used only within gcc.

> This whole thread on add-on libraries is way too pre-mature.
> gfortran can't pass the NIST F77 testsuite.  gfortran can't
> pass the LAPACK test programs (with netlib's BLAS).  There are
> 60+ nonstandard g77 intrinsic subprograms that need to be
> implemented for backwards compatibility.  At this point in
> time, we should be worried about making gfortran work.

I agree that you may not have time right now to worry about matrix-op 
performance; you guys are doing great work, and I'm not trying to be 
annoying.  And shipping add-on libraries in general is probably outside 
the scope of gcc.

However, when you do start to worry about matrix-op performance (and 
some people are already submitting patches), if you don't center your 
strategy around exploiting the BLAS then you are arguably making a mistake.

Steven