Re: [ublas] How to solve systems of linear equations?

Subject: Re: [ublas] How to solve systems of linear equations?
From: Karl Meerbergen (Karl.Meerbergen_at_[hidden])
Date: 2010-03-12 04:59:37

Rutger ter Borg wrote:
> Jesse Perla wrote:
>
>   
>> I was under the impression that most of the newer libraries took advantage
>> of vectorization for small, fixed size matrices through some mechanism,
>> but
>> I usually don't know what I am talking about.  ublas is a great,
>> stable,library and deserves a defense, even if its eventual deprecation is
>> considered.
>>     
>
> Indeed, that could be achieved through compile-time loop unrolling / mapping 
> that to specific compiler instructions. However, compilers have come a long 
> way, nowadays, it's nothing short but amazing what they're capable of, 
> including detection and unrolling of loops.
>   

I agree, the MTL4 team realized to their surprise that some of their 
template generated loop unrolled algorithms are actually worse than a 
compiled classical loop ...

Karl

>   
>> When will you have magma bindings for us :-)
>>
>>     
>
> By that time, it will be a matter of recompiling your lapack-bindings code 
> with something like BOOST_NUMERIC_BINDINGS_LAPACK_MAGMA (and/or containers 
> residing in GPU memory will trigger this automatically).
>
>   
>> I share the bias for linear algebra, but you still need to have a good
>> matrix library for basic operations (multiplication, slices, etc.) that is
>> highly expressive... so the container does matter a lot.
>>     
>
> You're right. That, and containers that support parallel/distributed memory 
> (to enable blacs/pblas/scalapack support), and gpu-memory stuff.
>
> Cheers,
>
> Rutger
>
>
>
> _______________________________________________
> ublas mailing list
> ublas_at_[hidden]
> http://lists.boost.org/mailman/listinfo.cgi/ublas
> Sent to: karl.meerbergen_at_[hidden]
>   

Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm