Re: [eigen] tri1 = tri2*tri1

[Gael Guennebaud <gael.guennebaud@xxxxxxxxx>]

indeed, for such problems, what you gain from skipping the known zeros is partly compensated by the extra logic to keep vectorization and more cache misses.

gael

On Tue, Sep 23, 2014 at 10:40 AM, Sebastian Birk <birk@xxxxxxxxxxxxxxxxxxxxx> wrote:

>> and currently I use dense square matrices for that with
>>   temp.tri = b*a.tri
>>   a.swap(temp)
> [...]
> I'm afraid there is nothing better at the moment. What are the typical
> sizes of your matrices?

They are rather small and range from 2 to about 100. So I guess there is
not much benefit in trying to optimise this.

I'm using them for block Krylov subspace methods where at some point QR
decompositions start to become the predominant computations.


Sebastian