Search found 90 matches

by hartwig anzt
Wed Nov 30, 2016 12:27 pm
Forum: User discussion
Topic: Preconditioner for cg
Replies: 1
Views: 1115

Re: Preconditioner for cg

Unfortunately, we currently have no AMG supported in MAGMA-sparse.

You may want to look into the ISAI preconditioner, this might be a faster alternative to an exact ILU:

./testing_dsolver --solver PCG --precond ILU --plevels 0 --trisolver ISAI --ppattern 1 --piters 0 testmatrix.mtx
by hartwig anzt
Tue Nov 08, 2016 11:51 am
Forum: User discussion
Topic: Magma change storage_type error in GPU
Replies: 4
Views: 1983

Re: Magma change storage_type error in GPU

Thank you for pointing this out! The cuSPARSE routine that is called in the converter converts the CSR matrix to a dense matrix stored in col-major. Col-major is also the LAPACK standard for dense matrices, so this is correct. Unfortunately, in MAGMA-sparse we always assume the dense matrices to hav...
by hartwig anzt
Thu Jun 23, 2016 8:18 am
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

Yes, we actually have some multi-GPU code that is not released. It looks like most people only use one GPU, maybe because most of the GPUs nowadays have a good amount of memory - the Tesla line typically 6-12 GB. Are the matrices confidential? Otherwise, I would appreciate if you could provide me wi...
by hartwig anzt
Wed Jun 22, 2016 4:09 pm
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

>> There are some sparse solvers I never heard about, though I am nowhere near a LA expert, like bombardment etc, so I might give those a try too! bombardment is not a Krylov solver by itself: it combines a number of Krylov solvers in interleaved fashion - i.e. QMR, CGS, BiCGSTAB. Thie idea is: if I...
by hartwig anzt
Wed Jun 22, 2016 8:52 am
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

Hey! I think the forum is the right place, as others may benefit from the discussion. And I truly welcome any feedback! 1) Yes, I do need to improve the documentation. I can tell you: available blocksizes are 2,4,8,16,32,64,128,256 available alignments are 1,4,8,16,32 but the combination may fail if...
by hartwig anzt
Tue Jun 21, 2016 10:31 am
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

Hey! Thanks for keeping me posted! Thanks for letting me know your experience! To answer your two questions: 1) This is what I would call a bug. Sorry! I will look into this - or just remove it as option, as the other one is faster anyway... 2) Although I don't want to stop you from looking into thi...
by hartwig anzt
Wed Jun 15, 2016 9:46 am
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

No - but I will keep this in mind for future. But you can always add a line in the tester for reading in a vector.
by hartwig anzt
Wed Jun 15, 2016 9:10 am
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

Thanks for the details! First of all, the shadow space dimension can be controlled via the parameter ''solver_par.restart''. So, there is no need to recompile the library, in fact you should not even have to touch the source files. In your runfile, if you want a shadow space dimension of 4, use zopt...
by hartwig anzt
Tue Jun 14, 2016 11:11 am
Forum: User discussion
Topic: IDR perfomance?
Replies: 14
Views: 5407

Re: IDR perfomance?

The IDR in AMGX sure is very optimized. However, I am surprised that the MAGMA version is that slow, in particular as it is identified to run close to the memory bandwidth: http://hpc.sagepub.com/content/early/2016/05/05/1094342016646844.abstract In particular, there should not be much GPU-host comm...
by hartwig anzt
Fri May 06, 2016 5:13 pm
Forum: User discussion
Topic: Sparse Solver Preconditioner
Replies: 2
Views: 1403

Re: Sparse Solver Preconditioner

Dear Li, indeed, it might happen that the cusparse ILU crashes in case of a zero diagonal entry. The implementation of the iterative ILU is according to the paper http://www.cc.gatech.edu/~echow/pubs/isc-2015.pdf It might also fail due to a zero diagonal, but you may give it a try by passing the arg...