[petsc-users] Fortran example for PETSc global to natural ordering test
Danyang Su
danyang.su at gmail.com
Mon Dec 3 19:32:44 CST 2018
Hi Matt,
Attached is the test example with source code, makefile, data and screen
output I wrote this afternoon. This example reads 2D mesh from vtk file
and then distribute over all processors. I can get correct global order
of local nodes after distribution, but the natural order of local nodes
is always zero after using DMPlexGlobalToNaturalBegin/End(). Sounds like
it does not take effect.
Would you mind helping me to check/test this code at you most convenience?
Thanks,
Danyang
On 2018-12-03 1:12 p.m., Danyang Su wrote:
>
>
> On 2018-12-03 12:56 p.m., Matthew Knepley wrote:
>> On Mon, Dec 3, 2018 at 3:40 PM Danyang Su <danyang.su at gmail.com
>> <mailto:danyang.su at gmail.com>> wrote:
>>
>>
>> On 2018-12-03 12:03 p.m., Matthew Knepley wrote:
>>> On Mon, Dec 3, 2018 at 2:27 PM Danyang Su <danyang.su at gmail.com
>>> <mailto:danyang.su at gmail.com>> wrote:
>>>
>>> Hi Matt,
>>>
>>> Thanks.
>>>
>>> BTW: DmPlexGetVertexNumbering now can work using the latest
>>> develop version. But the index is not in natural ordering
>>> when DMSetUseNatural is called. That's why I want to use
>>> PetscSFDistributeSection to check if I miss anything in the
>>> code.
>>>
>>> Can you explain that a little more? Maybe you can just push
>>> forward what you want using the migrationSF.
>>
>> Hi Matt,
>>
>> Since I cannot figure what is wrong or missing in my code, I
>> followed an old ex26.c example in
>> src/dm/impls/plex/examples/tests to create similar code as shown
>> below to test global to natural ordering. The code may be ugly
>> with unnecessary functions in it. Using DmPlexGetVertexNumbering,
>> I can get the value but it is not in natural order, instead, it
>> is still in default PETSc order without calling
>> DMSetUseNatural(dm,PETSC_TRUE,ierr).
>>
>> I do not understand what you are doing below. You just need to call
>>
>> ierr = DMSetUseNatural(dm,PETSC_TRUE);CHKERRQ(ierr);
>> ierr = DMPlexDistribute(dm,0,&migrationSF,&pdm);CHKERRQ(ierr);
>> if (pdm) {
>> ierr = DMPlexSetMigrationSF(pdm,migrationSF);CHKERRQ(ierr);
>> }
>> and the DMGlobalToNaturalBegin/End() should work.
>
> You mean to use DMPlexGlobalToNaturalBegin/End(), right? That's what I
> tried at first, but without success.
>
> I will create a test example to make further check if I can reproduce
> the problem.
>
> Thanks,
>
> Danyang
>
>>
>> Thanks,
>>
>> Matt
>>
>> if (rank == 0) then
>>
>> call
>> DMPlexCreateFromCellList(Petsc_Comm_World,ndim,num_cells,
>> num_nodes,num_nodes_per_cell, &
>> Petsc_False,dmplex_cells,ndim, dmplex_verts,dm,ierr)
>> CHKERRQ(ierr)
>> else
>> call
>> DMPlexCreateFromCellList(Petsc_Comm_World,ndim,0,
>> 0,num_nodes_per_cell, &
>> Petsc_False,dmplex_cells,ndim,dmplex_verts,dm,ierr)
>> CHKERRQ(ierr)
>> end if
>>
>> if (nprocs > 1) then
>> call DMSetUseNatural(dm,PETSC_TRUE,ierr)
>> CHKERRQ(ierr)
>> end if
>>
>> call DMPlexDistribute(dm,stencil_width, &
>> migrationsf,distributedMesh,ierr)
>> CHKERRQ(ierr)
>>
>> if (distributedMesh /= PETSC_NULL_DM) then
>> call
>> PetscSFCreateInverseSF(migrationsf,migrationsf_inv,ierr)
>> CHKERRQ(ierr)
>>
>> call
>> DMCreateGlobalToNatural(distributedMesh,migrationsf,migrationsf_inv,ierr)
>> CHKERRQ(ierr)
>>
>> call DMGetSection(distributedMesh,section,ierr)
>> CHKERRQ(ierr)
>>
>> call
>> PetscSectionCreate(Petsc_Comm_World,section_seq,ierr)
>> CHKERRQ(ierr)
>>
>> call PetscSFDistributeSection(migrationsf_inv,section, &
>> PETSC_NULL_INTEGER,section_seq,ierr)
>> CHKERRQ(ierr)
>>
>> call DMPlexCreateGlobalToNaturalSF(distributedMesh, &
>> section_seq,migrationsf,sf_natural,ierr)
>> CHKERRQ(ierr)
>>
>> call DMSetUseNatural(distributedMesh,PETSC_TRUE,ierr)
>> CHKERRQ(ierr)
>>
>> call PetscSFDestroy(migrationsf,ierr)
>> CHKERRQ(ierr)
>>
>> call PetscSFDestroy(migrationsf_inv,ierr)
>> CHKERRQ(ierr)
>>
>> end if
>>
>> Thanks,
>>
>> Danyang
>>
>>>
>>> Thanks,
>>>
>>> Matt
>>>
>>> Regards,
>>>
>>> Danyang
>>>
>>> On 2018-12-03 5:22 a.m., Matthew Knepley wrote:
>>>> I need to write a custom Fortran stub for this one. I will
>>>> get it done as soon as possible.
>>>>
>>>> Thanks,
>>>>
>>>> Matt
>>>>
>>>> On Sat, Dec 1, 2018 at 7:16 PM Danyang Su via petsc-users
>>>> <petsc-users at mcs.anl.gov <mailto:petsc-users at mcs.anl.gov>>
>>>> wrote:
>>>>
>>>> Hi All,
>>>>
>>>> I got a simple compilation error when use
>>>> PetscSFDistributeSection in
>>>> Fortran. It looks like the required head files are
>>>> included and the
>>>> parameters are correctly defined. However, when compile
>>>> the code, I got
>>>> error undefined reference to
>>>> `petscsfdistributesection_'. The code is
>>>> shown below. Did I miss anything here?
>>>>
>>>> #include <petsc/finclude/petscsys.h>
>>>> #include <petsc/finclude/petscvec.h>
>>>> #include <petsc/finclude/petscdm.h>
>>>> #include <petsc/finclude/petscdmplex.h>
>>>> use petscsys
>>>> use petscvec
>>>> use petscdm
>>>> use petscdmplex
>>>>
>>>> implicit none
>>>>
>>>> PetscSection :: section, section_seq
>>>> PetscSF :: migrationsf_inv, sf_natural
>>>> Vec :: vec_global, vec_natural
>>>> PetscErrorCode :: ierr
>>>>
>>>> ...
>>>>
>>>> call
>>>> PetscSFDistributeSection(migrationsf_inv,section,
>>>> &
>>>> PETSC_NULL_INTEGER,section_seq,ierr)
>>>> CHKERRQ(ierr)
>>>>
>>>>
>>>> call
>>>> PetscSFDistributeSection(migrationsf_inv,section, &
>>>> PETSC_NULL_INTEGER,section_seq,ierr)
>>>> CHKERRQ(ierr)
>>>>
>>>> Thanks,
>>>>
>>>> Danyang
>>>>
>>>>
>>>>
>>>> --
>>>> 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/>
>>
>>
>>
>> --
>> 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/>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20181203/24a677a8/attachment-0001.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: petsc-dmplex-test.tar.gz
Type: application/gzip
Size: 8518 bytes
Desc: not available
URL: <http://lists.mcs.anl.gov/pipermail/petsc-users/attachments/20181203/24a677a8/attachment-0001.bin>
More information about the petsc-users
mailing list