Hi Matt,
Thanks for the clarification, this does make sense. I have achieved the
local DoF ordering I want by only permuting the offsets in the created
section. However, I am now trying to extract the according halo exchange
information from the associated PetscSF object and, although the local
leafs have been renumbered according to my permutation, the roots they
are connected to have not been updated. Is there an easy way to update
the roots or re-initialise the PetscSF from the reordered section?
Many thanks
Michael
On 08/11/13 14:40, Matthew Knepley wrote:
On Thu, Nov 7, 2013 at 11:30 AM, Michael Lange
<[email protected] <mailto:[email protected]>>
wrote:
Hi,
I am trying to integrate an application code with DMPlex and I'm
hitting the following problem: The application code requires owned
and ghost DoFs to be numbered contiguously, whereas DMPlex numbers
them by entity class, so that the vertex dofs and edge dofs are
disjoint. In order to get around this problem I am trying to
re-order the DMPlex object vertically by using DMPlexPermute()
with a custom mapping such that the closure of cells is numbered
contiguously (except for repeated sieve points). This, however,
causes DMPlexCreateSection() to create wrong sections, because the
depth/height strata are not contiguous anymore; although I can
create the appropriate section from the original DMPlex and
permute it with the same mapping to get the correct section. From
this I am now trying to extract the appropriate information for my
halo exchange via the DefaultSF and the LocalToGlobalMap, and what
I was wondering is:
1) Is there an easy way to re-create the appropriate DefaultSF and
LocalToGlobalMap after the permutation, since these are ignored in
DMPlexPermute()?
2) Is it possible to use DMLabelGetStratumIS with the "depth"
label instead of DMPlexGetDepthStratum in DMPlexCreateSection() in
order to make this work with a vertically numbered DMPlex?
There is a misunderstanding here. You are confusing point numbering,
which is what is in DMPlex and what is modified by
DMPlexPermute, and dof numbering, which is handled by PetscSection.
The point numbering was originally arbitrary, which
made the code complex and slow. I don't think that the stratified
scheme will change.
However, I think all you need to interface with your legacy code is to
change the PetscSection ordering. This is fairly simple.
Right now, you set sizes and PetscSectionSetUp() calculated offsets,
but nothing in the code assumes that the offsets are
strictly ordered. Thus, you can just recalculate the offsets
traversing along your permuted ordering and calling
PetscSectionSetOffset().
Does this make sense?
Thanks,
Matt
Kind regards,
Michael Lange
--
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
