Re: [boost] uBLAS formal review

From: Martin Weiser (weiser_at_[hidden])
Date: 2002-06-28 04:20:37

On Freitag, 28. Juni 2002 00:24, Joerg Walter wrote:
> ----- Original Message -----
> From: "Martin Weiser" <weiser_at_[hidden]>
> > (i) Given the performance drop for large matrices, there should be a
> > possibility to supply high performance, specialized algorithms for
> > core algorithms such as _gemm. I tried to follow the instantiation
> > and execution path for C.assign(prod(A,B)) and got the impression
> > that it would be a hard job intercepting the expression template
> > chain in order to provide a specialized algorithm (correct me if I'm
> > wrong).
>
> We'd have to specialize vector_assign<> and matrix_assign<>.

Ok, if you say it's possible I'll be satisfied. After all you know the 
code better than me ;)

It's just that I considered specializing matrix_assign<> for dgemm and, 
reading the code, thought the necessary expression type information to be 
lost somewhere. Probably I'm just to dumb. Preferring a design that 
allows such specializations, I just wanted to draw your attention to this 
point.

Yours, 
Martin

-- 
Dr. Martin Weiser            Zuse Institute Berlin
weiser_at_[hidden]                Scientific Computing
http://www.zib.de/weiser     Numerical Analysis and Modelling