Re: [ublas] OpenMP support

Subject: Re: [ublas] OpenMP support
From: Oswin Krause (Oswin.Krause_at_[hidden])
Date: 2011-11-17 18:03:27

Hi,

i agree. There is no need to even think about OpenMP and parallelization
stuff as long as single core implementations of, for example, the dense
matrix-matrix product of MKL can outperform ublas by more than an order of
magnitude.

I don't think that more functions like axpy_prod are the solution. They
undermine the goal of getting rid of the old c-style interfaces.
Unfortunately, i doubt that it is possible to get the faster algorithms
working with the current way of evaluating expressions.

At the moment i would vote for: rewrite ublas such that it makes use of as
much of the numeric bindings as possible with a reasonable fall-back in
the case that no BLAS library is used. This avoids the problem of
implementing the more elaborate algorithms.

In any case, a lot of code using ublas will break.

Just my 2 cents :)

Greetings,
Oswin

> no and yes.
> So far we have no parallelization at all, apart from the implicit
> vectorization made by the compiler. :-(
> Yes it's in the pipe but honnestly, ublas needs to be transformed a little
> bit. I had a look at how Eigen is doing all the vector/parallel/multicore
> things and we need to make some change into the code base before doing
> that.
>
> Also have a look at the numeric bindings in the boost repository, they
> give
> you access to plenty of nice fast code.
>
> David
>
>
> On Thu, Nov 17, 2011 at 12:04, Kraus Philipp
> <philipp.kraus_at_[hidden]>wrote:
>
>> Hello,
>>
>> I'm using a lot of Boost UBlas structures. Does the UBlas structures use
>> OpenMP? Is there a support in the future?
>>
>> Thanks
>>
>> Phil
>> _______________________________________________
>> ublas mailing list
>> ublas_at_[hidden]
>> http://lists.boost.org/mailman/listinfo.cgi/ublas
>> Sent to: david.bellot_at_[hidden]
>>
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: Oswin.Krause_at_[hidden]