Hi Matthew, I think I need the whole G2L not just the sizes. Build my own matrix, create a “fake” DMDA, extract the G2L mapping and apply it to my preallocation/assembly routines (where I would basically replace my natural ordering with the DMDA ordering from the G2L mapping)
(cf the mail I sent to the user list approx. 1 hour ago) Thibaut On 22 Feb 2019, at 14:24, Matthew Knepley <knep...@gmail.com<mailto:knep...@gmail.com>> wrote: On Thu, Feb 21, 2019 at 1:19 PM Thibaut Appel <t.appe...@imperial.ac.uk<mailto:t.appe...@imperial.ac.uk>> wrote: Hi Matthew, Is your first part of your answer (using DMDASetBlockFills) valid only in the case I create a DMDA object? Yes I think that is the kind of stencil I am using. I could know how the stencil looks like exactly, but I preallocate looping, for each process on all the elements of the stencil, grid node by grid node (which is not that costly, and "exact") If I do NOT use a DMDA object and create my MPIAIJ matrix myself, how do I get the global row indices owned by the process (the "DMDA-like" ones you mentioned)? I think you want https://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/DMDA/DMDAGetLocalInfo.html This tells you exactly what patch is owned by the process, and also how big the overlap is. Its not the same as G2L, but do you need the whole G2L, or just the sizes? Thanks, Matt The problem is that MatGetOwnershipRange cannot be called if the matrix hasn't been preallocated, and I need the global indices to preallocate. Thibaut On 21/02/2019 17:49, Matthew Knepley wrote: On Thu, Feb 21, 2019 at 11:16 AM Thibaut Appel via petsc-users <petsc-users@mcs.anl.gov<mailto:petsc-users@mcs.anl.gov>> wrote: Dear PETSc developers/users, I’m solving linear PDEs on a regular grid with high-order finite differences, assembling an MPIAIJ matrix to solve linear systems or eigenvalue problems. I’ve been using vertex major, natural ordering for the parallelism with PetscSplitOwnership (yielding rectangular slices of the physical domain) and wanted to move to DMDA to have a more square-ish domain decomposition and minimize communication between processes. However, my application is memory critical, and I have finely-tuned matrix preallocation routines for allocating memory “optimally”. It seems the memory of a DMDA matrix is allocated along the value of the stencil width of DMDACreate and the manual says about it “These DMDA stencils have nothing directly to do with any finite difference stencils one might chose to use for a discretization” And despite reading the manual pages there must be something I do not understand in the DM topology, what is that "stencil width" for then? I will not use ghost values for my FD-method, right? What this is saying is, "You might be using some stencil that is not STAR or BOX, but we are preallocating according to one of those". If you really care about how much memory is preallocated, which it seems you do, then you might be able to use https://www.mcs.anl.gov/petsc/petsc-current/docs/manualpages/DMDA/DMDASetBlockFills.html to tell use exactly how to preallocate. I was then wondering if I could just create a MPIAIJ matrix, and with a PETSc routine get the global indices of the domain for each process: in other words, an equivalent of PetscSplitOwnership that gives me the DMDA unknown ordering. So I can feed and loop on that in my preallocation and assembly routines. You can make an MPIAIJ matrix yourself of course. It should have the same division of rows as the DMDA division of dofs. Also, MatSetValuesStencil() will not work for a custom matrix. Thanks, Matt Thanks very much, Thibaut -- 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 https://www.cse.buffalo.edu/~knepley/<http://www.cse.buffalo.edu/~knepley/> -- 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 https://www.cse.buffalo.edu/~knepley/<http://www.cse.buffalo.edu/~knepley/>
