Home My Page Projects PaStiX
Summary Activity Forums Lists Docs News Files

Forum: open-discussion

Monitor Forum | Start New Thread Start New Thread
RE: Using IPARM_DOF_NBR [ Reply ]
By: Xavier Lacoste on 2013-03-20 10:11
[forum:110793]
Are you using the distributed interface of PaStiX (with l2g) ?

I ran into a bug when IPARM_DOF_NBR > 1, the matrix is distributed and the distribution doesn't match the solver distribution.

In that particular case we are redistributing the matrix without taking into account the number of degree of freedom.

So if you are in that case, I'll send you a snpashot with a fix.

XL.

RE: Using IPARM_DOF_NBR [ Reply ]
By: Xavier Lacoste on 2013-03-12 07:53
[forum:110718]
Yes it's used on some codes I know.

This information is missing in the ref card, you are right.

The CSC is a blocked CSC, like a normal CSC but you have ndof*ndof values by node, in a column major mode.

eg :

0 1 . . . .
2 3 . . . .
. . 4 5 . .
. . 6 7 . .
8 9 . . 10 11
12 13 . . 14 15

with ndof = 2 would give :
n = 3
colptr = {1, 3, 4, 5}
rows = {1, 3, 2, 3}
values = { 0, 2, 1, 3, 8, 12, 9, 13, 4, 6, 5, 7, 10, 14, 11, 15}

Is this explanation sufficient ?

XL.

RE: Using IPARM_DOF_NBR [ Reply ]
By: Garth Wells on 2013-03-11 18:43
[forum:110716]
Sorry it's been a while, but I'm returning to this with PaStiX 4030.

Has IPARM_DOF_NBR > 1 been tested? I only get seg faults.

Also I can't find in the documentation the sparse matrix format that PaStiX expects when IPARM_DOF_NBR > 1. Is it documented somewhere?

Is it block compressed sparse column or still just compressed sparse column?

Garth

RE: Using IPARM_DOF_NBR [ Reply ]
By: Xavier Lacoste on 2012-07-05 06:43
[forum:109839]
The murge.c/fmurge.F90 examples build a laplacian with IPARM_DOF_NBR >= 1.
The murge-product.c also works with DOF >= 1.

I don't think we have an example with the non-murge interface using DOF > 1 because they are all based on the same matrix reading code which doesn't work with DOF > 1.

Sorry for the segfaults...
I hope we can fix that by adding some more understandable error message...
Can I have your code to reproduce it ?

Thanks.

XL.

EDIT: And for the performance, it should improve the pre-processing steps in term of time and memory as the graph is smaller.

Using IPARM_DOF_NBR [ Reply ]
By: Garth Wells on 2012-07-04 23:11
[forum:109837]
Is there example code somewhere that uses IPARM_DOF_NBR > 1? I've tried using block compresses matrix, and the matrix check passes but PaStiX segfaults when doing the re-ordering.

Is using IPARM_DOF_NBR > 1 for vector problems likely to give much of a performance improvement?

PaStiX does segfault a lot when trying to get it working. Some assertions in debug mode would really help users.

Garth