Re: [boost] Re: numeric/ublas ?

From: Toon Knapen (toon.knapen_at_[hidden])
Date: 2001-11-13 06:31:39

walter_at_[hidden] wrote:

> --- In boost_at_y..., Toon Knapen <toon.knapen_at_s...> wrote:
> 
> ublas is somewhat scattered in the CVS (branch matrix_development). 
> You'll find the odds at:

[snip]

but e.g. the includes in the header of test1 refer to "../config.h" 
which seems incompatible with the directory structure you use in boost 
IMO : There is no subdirectory in the directory where config.h is 
located ? (I figure you use some other file-structure layout to compile 
the tests?).
Anyway, I'm very interested in using uBlas and could also produce 
Jamfiles if you like to build the tests automatically (I would suggest 
moving the ublas on the main branch in that case however as otherwise 
synchronising with the latest Jam developments might be difficult)

> 
>>And what is the current status of ublas ? 
>>
>  
> Rather advanced, w.r.t. Fortran BLAS functionality, I believe ;-). We 
> are still writing docs and try to tune the kernels...

What is the relation to MTL3.0 development in Boost ?

> 
>>Do you use Blas internally for 
>>performance reasons (I checked but could no see any calls to the 
>>
> blas.h 
> 
>>header)
>>
>  
> Nope.The template functions in blas.h are used to demonstrate how to 
> emulate Fortran BLAS (see test2). However, if one needs platform 
> specific tuned BLAS kernels, one could try to partially specialize 
> these.

I would like to look at making specialisations for using Blas internally
(too have the most performant alternative as long as the C++ 
implementation is at par with Blas) if it does not collide with the 
roadmap you have in mind.

toon