[ADMB Users] LU decomposition revisited

[H. Skaug]

Great. In the context of spatial statistics "2000x2000 " means that you
can split your regioin into 2000 sub-areas where you have observations.

Parallization: I think everyone agrees that parallization is crucial for
the future of ADMB, although the amount of amount feedback you have received
on your earlier emails have not reflected this. For sparse matrix
calculations I hope to be able to utilize

http://www.cise.ufl.edu/research/sparse/

I am not sure how applicable this is to dense matrices, but it
supports GPU calculations.
It would be nice to have a unified approach to parallization, but I am not sure
if that is realistic.

Hans



On Sun, Nov 25, 2012 at 8:09 PM, dave fournier <davef at otter-rsch.com> wrote:
> For big spatial random effects models one needs to do the LU decomposition
> on
> large matrices.  (could use choleski I guess but I wanted to revisit the
> LU).
>
> I recall that there were some performance issues with this function so there
> was
> a rollback.  (Maybe someone can refresh my memory on this issue) Anyway I
> was surprised
> because the new version was supposed to access memory more efficiently.
> Maybe the
> problem was with the variable (dvar_matrix) version. In any event I set up a
> little test
> for comparison (attached)  For a 2000x2000 matrix the newer code is already
> 6 times faster.  It also has a good structure for mult-threading. Anyway I
> think we can make some
> good gains here.