Boost logo

Ublas :

Subject: Re: [ublas] sparse matrix dense matrix multiplication dilemma
From: Umut Tabak (u.tabak_at_[hidden])
Date: 2011-09-26 02:29:52


On 09/26/2011 08:12 AM, Ungermann, Jörn wrote:
> Dear Umut,
>
> This is a problem. It depends a bit on the type of the target matrix. I
> suppose that to be dense, too.
> I proposed a fast product routine last year, but nobody seems to be doing
> something about it:
> http://lists.boost.org/MailArchives/ublas/2010/03/4091.php
> You may have a look at that mail, which contains a product routine, which is
> reasonably fast for most involved matrix types. If it isn't, it's probably
> not possible to be fast without a priori information.
> The proposal contains a series of rather general multiplication kernels,
> which are automatically chosen according to the involved types. If you have
> problems checking out that code, please send me a direct message.
>
> According to my results from the cited mail, your best bet within
> boost::ublas for sparse_row-major_matrix times dense_row-major_matrix to
> dense_row-major_matrix would be "sparse_prod".
>
> Before you switch to gm++, please check out also eigen3 and tell me your
> experiences after an eventual switch.
>
> Regards,
> Joern
>
>
Thanks for the comments, where is 'sparse_prod' defined, I mean in which
header?
I will take a look at eigen3 however their sparse matrices were also
under major revision last time I checked.
That was the reason to have a look at gmm++ which seems reasonably faster.
Best,
Umut?