[eigen] Re: cache-friendly matrix inverse

[ Thread Index | Date Index | More lists.tuxfamily.org/eigen Archives ]


ah, and i was a bit pessimistic in my "relative error" estimation:
the norm() of the identity matrix is actually sqrt(size), not 1, since
we're talkign about the l2 norm here. So actually, the precision is
35x better than i measured. Still, that doesn't change much -- still
unusable for large float matrices.

Benoit

2009/5/13, Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
> 2009/5/13, Benoit Jacob <jacob.benoit.1@xxxxxxxxx>:
>> it spends most of its time in cache friendly matrix product
>
> oops, this is inexact.
>
> i wanted to say that i tried reducing all the way to tiny fixed-size
> blocks, and then it spend most of its time in cache-friendly product
> indicating that i couldn't squeeze any better performance out of this
> approach; still the performance wasn't better than what i attached to
> the previous email where it calls partial LU when blocks are small
> enough.
>
> Benoit
>



Mail converted by MHonArc 2.6.19+ http://listengine.tuxfamily.org/