Re: [eigen] A vector indexed by another vector

[Christoph Hertzberg <chtz@xxxxxxxxxxxxxxxxxxxxxxxx>]

On 24.08.2012 16:31, Matthieu Brucher wrote:
> > This has been requested several times, essentially this is directly what
> > bug 329 wants to have:
> > http://eigen.tuxfamily.org/bz/**show_bug.cgi?id=329<http://eigen.tuxfamily.org/bz/show_bug.cgi?id=329>
> > There is a prototype implementation, I'm not sure how well it worked/works
> > with the current Eigen version.
> >
> > If, however, the offset array is a permutation, i.e. each index occurs
> > exactly once, you should be able to solve it using:
> > http://eigen.tuxfamily.org/**dox-devel/classEigen_1_**
> > 1PermutationMatrix.html<http://eigen.tuxfamily.org/dox-devel/classEigen_1_1PermutationMatrix.html>
>
> Thanks for the pointers. Unfortunately, I know that there are indices that
> will not be used. We do skip some values depending on a parameter at the
> beginning of the process, so perhaps duplicating data would be an
> acceptable solution (although a memory- and time-costly one, there are
> dozens of arrays that may or may not be needed).

No, duplicating the data definitely is not an option, both for reasons 
of performance and elegance-of-code.

The cleanest solution, obviously, would be to fix bug 329 ;).
Until you or someone else does so, I would recommend keeping the old code.

If you really need to go for performance, it might be an option to 
extract all data which is indexed via offset[j] manually into 
temporaries. That way the remaining expression could even benefit from 
SSE-vectorization. However, this obviously might be memory-costly.



Christoph


--
----------------------------------------------
Dipl.-Inf. Christoph Hertzberg
Cartesium 0.049
Universität Bremen
Enrique-Schmidt-Straße 5
28359 Bremen

Tel: +49 (421) 218-64252
----------------------------------------------