Re: [ublas] Matrix multiplication performance

Subject: Re: [ublas] Matrix multiplication performance
From: Riccardo Rossi (rrossi_at_[hidden])
Date: 2016-01-28 17:02:48

Well there is pastix...which is c99 and performs way better than superlu.

Now seriously I think ublas misses a sparse spmm algorithm.

The fastest implementation I am aware of is in viennacl. (C++)

The limitation for writing a good one in ublas is mostly that it misses a
constructor in which csr arrays are passed already built (ideally via move
semantics I guess).

If you redesign ublas think about how to provide such access!

Cheers
Riccardo
On 28 Jan 2016 22:01, "Karl Meerbergen" <karl.meerbergen_at_[hidden]>
wrote:

>
> > On 28 Jan 2016, at 21:47, Michael Lehn <michael.lehn_at_[hidden]> wrote:
> >
> >
> >
> >> do you also do sparse linear algebra by chance?
> >
> > Sorry, not directly.  I just looked at libraries like SuperLU and
> Umfpack.  However, not as close as to other BLAS libraries.  But
> > from my impression this also could be done much more elegant in C++.
> The big headache in these libraries is that they basically
> > have the same code for float, double, complex<float> and complex<double>
> .  Just using C++ as "C plus function templates” would
> > make it much easier.  And the performance relevant part in these
> libraries is again a fast dense BLAS.
>
> Correct, but I would bet on MUMPS, which is,in my opinion, more advanced
> and still improving. They also use a ’template’ mechanism in fortran 90,
> based on the C preprocessor ;-) They made it clear they will not redo their
> developments of more than 30 man years in C++.
>
> Best,
>
> Karl
>
> Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: rrossi_at_[hidden]
>