search

ecrc / hicma

HiCMA: Hierarchical Computations on Manycore Architectures
https://ecrc.github.io/hicma/
BSD 3-Clause "New" or "Revised" License
27 stars 11 forks source link

I'm using hicma to do some tests. I have just one question #1

Open joecomp opened 4 years ago

joecomp commented 4 years ago

Hi

I'm very interested by the HICMA software which seems very good.

Do you plan to add soon some other features as plaaned in your presentation ?

I'm especially interested by the LU Factorization / Solve , the OPENMP support and the support of other hierarchical format (H format in first)

Thank you very much in advance for your answer.

Regards

ltaiefhatem commented 4 years ago

Hi, Thanks for your interests! We have just added support for LU factorization and solve on shared and (soon) distributed-memory systems. We have been also working on integrating al4san (https://github.com/ecrc/al4san) to abstract the dynamic task-based runtime systems, which currently supports StarPU, PaRSEC, Quark, and OpenMP runtime systems. The new HiCMA code will be released in a few weeks, once the extensive testing and benchmarking phases successfully terminate on various hardware architectures. As far as the compression format, we are looking into extending to HODLR for now. Stay tune! Hatem

joecomp commented 4 years ago

Dear Hatem

I'm doing some tests with HICMA library that is promising for me.

I have just one small additional question.

Is it possible to use complex number with cholesky factorization ? The example testing_ztrsmd seems to be with real matrices.

Thanks in advance for your reply.

Best regards

Johnny

ltaiefhatem commented 4 years ago

Dear Johnny, Absolutely. This should be doable. @kadircs and I will look into it and come back to you shortly. Thanks, H

joecomp commented 4 years ago

Dear Hatem

Thanks for your answer.

What I use is complex matrices (real and imaginary part) providing from acoustic field so I was wondering if it was possible to use such kind of matrices with HICMA.

Have a great day or great evening

Regards

ltaiefhatem commented 4 years ago

Dear Johnny, If your matrix is hermitian positive definite, we should be able to provide you the complex version of the current release of Cholesky. We will be glad to have your feedback and if you could also test it on your side. Thanks, H

joecomp commented 4 years ago

Dear Hatem

my matrix is complex symmetric positive not hermitian just symmetric (I use the lapack function ZSYTRF for now)

If you can send me a version for this kind of matrix it will be wonderful and I will be happy to test it.

However if you have just a complex version with equivalent to zpotrf lapack function I will be interested to have it.

Thanks again for your help and support

Regards

kadircs commented 4 years ago

Dear @joecomp, Did you try to do SVD on your complex matrix? How are the ranks when you select singular values larger than an arbitrary threshold (e.g., 1e-3)? I would like to request a piece of code to generate your complex positive-definite matrix so that I can directly test complex HiCMA_zpotrf() on your matrix.

joecomp commented 4 years ago

Dear Kadir

Sorry to answer you so late , it's difficult for me to give you a piece of code to generate the matrix (it's too complex).

What I can give you is a small example of the kind of matrix I have in text format (please find attached it).

The file contains the upper triangular part of the matrix done with 614 equations (matrix contains 614 * 615 / 2 = 188 805 terms in total ) and is stored by rows.

First 614 complex values is the first row of the matrix , then the next 613 values is the second row of the matrix , etc...

Is this sufficient for you to do your test or do you want something else ?

Thenk you very much for your support and your time.

Best Regards DBEM.zip

joecomp commented 4 years ago

Dear Hatem , dear Kadir

I hope all is well with you and your family in these difficult times

Do you have an idea on when the new version of your software will be available ?

I have seen the abstract of your paper for the ISC conference "Solving Acoustic Boundary Integral Equations Using High Performance Tile Low-Rank LU Factorization".

I will be interested to have the full paper if it possible for you to send me it.

Thank you very much for your answer and your support.

Have a good day.

Regards

kadircs commented 4 years ago

Dear Johnny, The LU factorization for double complex precision is available here: https://github.com/ecrc/hicma/tree/zgetrf Would you please ask your questions about this branch in a separate issue? This issue #1 (zpotrf) is currently being addressed internally. I am sorry for the delay. I will also use your data (DBEM.zip) for testing zpotrf. Thanks, Kadir