Boost logo

Boost :

From: lums_at_[hidden]
Date: 2001-03-26 20:16:37


--- In boost_at_y..., boost_at_s... wrote:
> Salut
>
> [...]
> --- In boost_at_y..., lums_at_l... wrote:
> > --- In boost_at_y..., boost_at_s... wrote:
> > > the left, and B_i appears on the right.
> > > So using different storage layouts, I can increase performance
> > > since no strides appear.
> >
> > For dense time dense the storage layout of the original matrices
> is
> > somewhat irrelevant for performance issues. By performing
> > appropriate blocking and re-arranging for optimal layout within
> the
> > working blocks, one can get optimal performance for any original
> > layout. ATLAS, PhiPAC, optimized BLAS, and MTL all do this AFAIK.
> >
> > That is, yes, layout is important, but only in the working blocks.
>
> > Any original storage format can be copied to a good orientation in
> > the working blocks within the mat-mult algorithm.
>
> O.k., I never written blocked versions of matmuls, so I assume
> you're right.
>
> Nevertheless one should provide both memory layouts, allthough
> one may name it differently, for representing transposed matrices.

Yes, absolutely, both kinds of orientation should be supported (as
well as diagonal). The point was that there was no need to pick a
single layout. Or that may have been the point. We're getting kind
of deep in the message nesting....


Boost list run by bdawes at acm.org, gregod at cs.rpi.edu, cpdaniel at pacbell.net, john at johnmaddock.co.uk