Dense and Sparse LU decomposition

Open discussion for MAGMA library (Matrix Algebra on GPU and Multicore Architectures)
Post Reply
Linuxboy
Posts: 15
Joined: Tue Nov 29, 2011 9:24 pm

Dense and Sparse LU decomposition

Post by Linuxboy » Mon Oct 13, 2014 4:30 pm

Good day everybody! I've got a question. A have a system with sparse matrix NxN, then I do some transformations and got sparse matrix 2Nx2N. Then solve system with LU like dense (i got enough GPU memory). If i use somedirect/iiterative solver is it be slower or faster? Where i can find some results of the comparison?
In my current work i transform dense matrix of system in some kind of sparse matrix (relative sparseness 0,5). Then use LU decomp for solving. I've got great performance on Nvidia Titan!

mgates3
Posts: 897
Joined: Fri Jan 06, 2012 2:13 pm

Re: Dense and Sparse LU decomposition

Post by mgates3 » Mon Oct 13, 2014 5:21 pm

The speed of sparse solvers is highly dependent on the matrix itself, e.g., its sparsity pattern, convergence rate with CG or GMRES, applicable preconditioners, etc. Try out the sparse solvers in
magma/sparse-iter/testing/

-mark

Linuxboy
Posts: 15
Joined: Tue Nov 29, 2011 9:24 pm

Re: Dense and Sparse LU decomposition

Post by Linuxboy » Tue Oct 14, 2014 11:43 am

Hi mark!
In what format should be a file for the function magma_d_vread? I try to read vector - i got error. In functions like dgetrf i read this file ok/

hartwig anzt
Posts: 90
Joined: Tue Sep 02, 2014 5:44 pm

Re: Dense and Sparse LU decomposition

Post by hartwig anzt » Fri Oct 17, 2014 11:29 am

Dear Linuxboy,

thank you for sharing your observations. I am aware of this problem, and already fixed it - the bugfix will be included in the next release.

If you need this function really urgent, please let me know, I will send you the updated file.

Thanks, Hartwig

Post Reply