date:20190319

Re: [petsc-users] Saving Vecs/Mats in HDF5 and visualizing in Matlab

2019-03-19 Thread Yuyun Yang via petsc-users

Sounds good, thanks for the advice!

-Original Message-
From: Jed Brown  
Sent: Tuesday, March 19, 2019 1:23 PM
To: Yuyun Yang ; zakaryah ; 
petsc-users@mcs.anl.gov
Subject: Re: [petsc-users] Saving Vecs/Mats in HDF5 and visualizing in Matlab

Yuyun Yang via petsc-users  writes:

> It's simply for visualization purposes. I wasn't sure if HDF5 would perform 
> better than binary, and what specific functions are needed to load the PETSc 
> vectors/matrices, so wanted to ask for some advice here. Since Matt mentioned 
> it's not likely to be much faster than binary, I guess there is no need to 
> make the change?

There is no speed benefit.  The advantage of HDF5 is that it supports better 
metadata, including the data types and sizes.  The PETSc data format is quick, 
dirty, and fast.

Re: [petsc-users] MatCompositeMerge + MatCreateRedundantMatrix

2019-03-19 Thread Marius Buerkle via petsc-users

If it is not too complicated it would be nice if MPICreateSubMatricesMPI would be avaible for other matrices than MPIAIJ.

On Mon, Mar 18, 2019 at 10:30 PM Marius Buerkle wrote:

I have another question regarding MatCreateRedundantMatrix and MPICreateSubMatricesMPI. The former works for MPIAIJ and MPIDENSE and the later only for MPIAIJ. Would it be possible to use MatCreateRedundantMatrix with a factored matrix

We usually do not have direct access to the data for factored matrices since it lives in the factorization package.

and MPICreateSubMatricesMPI with dense and/or elemental matrices ?

This would not be hard, it would just take time. The dense case is easy. I think Elemental already has such an operation, but we would have to find it and call it.

Thanks,

Matt

Indeed, was very easy to add. Are you going to include the Fortran interface for MPICreateSubMatricesMPI in future releases of PETSC ?

Regarding my initial problem, thanks a lot. It works very well with MPICreateSubMatricesMPI and the solution can be implemented in a few lines.

Thanks and Best,

Marius

On Tue, Mar 12, 2019 at 4:50 AM Marius Buerkle wrote:

I tried to follow your suggestions but it seems there is no MPICreateSubMatricesMPI for Fortran. Is this correct?

We just have to write the binding. Its almost identical to MatCreateSubMatrices() in src/mat/interface/ftn-custom/zmatrixf.c

Matt

On Wed, Feb 20, 2019 at 6:57 PM Marius Buerkle wrote:

ok, I think I understand now. I will give it a try and if there is some trouble comeback to you. thanks.

Cool.

Matt

marius

On Tue, Feb 19, 2019 at 8:42 PM Marius Buerkle wrote:

ok, so it seems there is no straight forward way to transfer data between PETSc matrices on different subcomms. Probably doing it by "hand" extracting the matricies on the subcomms create a MPI_INTERCOMM transfering the data to PETSC_COMM_WORLD and assembling them in a new PETSc matrix would be possible, right?

That sounds too complicated. Why not just reverse MPICreateSubMatricesMPI()? Meaning make it collective on the whole big communicator, so that you can swap out all the subcommunicator for the aggregation call, just like we do in that function.

Then its really just a matter of reversing the communication call.

Matt

On Tue, Feb 19, 2019 at 7:12 PM Marius Buerkle wrote:

I see. This would work if the matrices are on different subcommumicators ? Is it possible to add this functionality ?

Hmm, no. That is specialized to serial matrices. You need the inverse of MatCreateSubMatricesMPI().

Thanks,

Matt

marius

You basically need the inverse of MatCreateSubmatrices(). I do not think we have that right now, but it could probably be done without too much trouble by looking at that code.

Thanks,

Matt

On Tue, Feb 19, 2019 at 6:15 AM Marius Buerkle via petsc-users wrote:

Hi !

Is there some way to combine MatCompositeMerge with MatCreateRedundantMatrix? I basically want to create copies of a matrix from PETSC_COMM_WORLD to subcommunicators, do some work on each subcommunicator and than gather the results back to PETSC_COMM_WORLD, namely I want to sum the the invidual matrices from the subcommunicatos component wise and get the resulting matrix on PETSC_COMM_WORLD. Is this somehow possible without going through all the hassle of using MPI directly?

marius

What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead.
-- Norbert Wiener