Re: [boost] Re: uBLAS and expression templates

From: Joerg Walter (jhr.walter_at_[hidden])
Date: 2002-07-01 23:12:30

----- Original Message -----
From: "David Abrahams" <david.abrahams_at_[hidden]>
To: <boost_at_[hidden]>
Sent: Monday, July 01, 2002 3:28 PM
Subject: Re: [boost] Re: uBLAS and expression templates

>
> ----- Original Message -----
> From: "Joerg Walter" <jhr.walter_at_[hidden]>
>
> > A couple of questions remain open:
> > - Should we rename the original ET' based prod() functions to something
> like
> > defer_prod()? - Is it possible to determine a sensible default for the
> type
> > of the temporary (for example via some traits classes)?
>
> I'm not familiar enough with the implementation of uBlas to understand
> what's happening here. However, I think I proved in my recent posting that
> no temporaries are neccessary, and that you might want to *consider* using
> them for A*(B*v) to reduce the number of writes. It wasn't clear to me
> how/whether you were responding to that post, though...

I think, I've understood, that you propose to evaluate A * (B * v) via

       r.clear ();
        for (unsigned j = 0; j < r.size (); ++ j)
            r.plus_assign (column (A, j) * prod (B, v) (j));

May be we even could defer the loop evaluation via some lambda construct.
I've two problems with this approach: I'm not sure if the results of the
different evaluation path couldn't surprise a client of ublas. And I
currently don't know, how this optimization generalizes to matrix chain
multiplication and other expressions.

Regards

Joerg