Re: matrix library ?

From: Toon Knapen (toon_at_[hidden])
Date: 2001-03-12 08:51:05

> 
> 
> Hello, 
>     TK> Would there be any interest in developing a multi-dimensional
>     TK> array (vector, matrix, ...) library at Boost ?
> 
> There have actually been some proposals for such a library (there's an 
> old boost thread headed by Giovanni Ballestri at the end of last
> year).  I've been using ideas presented in his work and the discussion 
> that evolved around it to develop a library with very similar intent.
> It hasn't yet proceeded to the point where I'm ready to propose it to
> boost, but I am definitely interested in conversation about the topic.
>
> ron

I've been looking at the the multi_array in the vault and was thinking
if it could be used as a basis for a multi-array lib in boost.

As far as I'm concerned, and keeping in mind the previous discussions 
about this topic, the vision/goals of the library should be
1_ optimal performance (on par with fortran or better)
2_ user friendly interface (~f90 or matlab)
3_ easy to interface with other libraries

I think the first requirement can only be satisfied if
cache-optimisation
by means of blocked operations would be included. When we want to
support
fixed size (small) arrays to be contained in dynamic sized arrays, we
can't
reuse the existing multi_array.zip though.

I've included the third requirement because it might be possible on some
platforms that there are vendor-tuned libraries (usually BLAS) present
with a performance that can not be obtained by a multi-platform lib
easily.
Thus if, depending on the configuration, the multi-array library would
sometimes make use of these vendor tuned libraries without changing
something
to the interface of the arrays (and thus not visible to the lib-user),
we would have the best of both worlds.

Thus should we start from scratch or ... ?

toon