[petsc-users] Re-ordering in DMPlexCreateFromCellListParallelPetsc
Matthew Knepley
knepley at gmail.com
Mon Mar 22 11:53:06 CDT 2021
On Mon, Mar 22, 2021 at 12:20 PM Nicolas Barral <
nicolas.barral at math.u-bordeaux.fr> wrote:
> Thanks for your answers Matt.
>
> On 22/03/2021 13:45, Matthew Knepley wrote:
> > On Mon, Mar 22, 2021 at 6:22 AM Nicolas Barral
> > <nicolas.barral at math.u-bordeaux.fr
> > <mailto:nicolas.barral at math.u-bordeaux.fr>> wrote:
> >
> > On 21/03/2021 21:29, Matthew Knepley wrote:
> > > On Sat, Mar 20, 2021 at 10:07 AM Nicolas Barral
> > > <nicolas.barral at math.u-bordeaux.fr
> > <mailto:nicolas.barral at math.u-bordeaux.fr>
> > > <mailto:nicolas.barral at math.u-bordeaux.fr
> > <mailto:nicolas.barral at math.u-bordeaux.fr>>> wrote:
> > >
> > > Hi all,
> > >
> > > I'm building a plex from elements arrays using
> > > DMPlexCreateFromCellListParallelPetsc. Once the plex is
> built, I
> > > need to
> > > set up boundary labels. I have an array of faces containing a
> > series of
> > > 3 vertex local indices. To rebuild boundary labels, I need to
> > loop over
> > > the array and get the join of 3 consecutive points to find the
> > > corresponding face point in the DAG.
> > >
> > >
> > > This is very common. We should have a built-in thing that does
> this.
> > >
> > Ordering apart, it's not very complicated once you figure out "join"
> is
> > the right operation to use. We need more doc on graph layout and
> > operations in DMPlex, I think I'm going to make pictures when I'm
> done
> > with that code because I waste too much time every time. Is there a
> > starting point you like ?
> >
> >
> > I think the discussion in
> >
> > @article{LangeMitchellKnepleyGorman2015,
> > title = {Efficient mesh management in {Firedrake} using
> > {PETSc-DMPlex}},
> > author = {Michael Lange and Lawrence Mitchell and Matthew G.
> > Knepley and Gerard J. Gorman},
> > journal = {SIAM Journal on Scientific Computing},
> > volume = {38},
> > number = {5},
> > pages = {S143--S155},
> > eprint = {http://arxiv.org/abs/1506.07749},
> > doi = {10.1137/15M1026092},
> > year = {2016}
> > }
> >
> > is pretty good.
> >
> It is, I'd just like to have something more complete (meet, join,
> height, depth...) and with more 2D & 3D pictures. It's all information
> available somewhere, but I would find it convenient to be all at the
> same place. Are sources of the paper available somewhere ?
>
I think the right answer is no, there is not a complete reference. I have
some things in presentations,
but no complete work has been written. Maybe it is time to do that. At this
point, a book is probably
possible :)
> > > Problem, vertices get reordered by
> > > DMPlexCreateFromCellListParallelPetsc
> > > so that locally owned vertices are before remote ones, so
> > local indices
> > > are changed and the indices in the face array are not good
> > anymore.
> > >
> > >
> > > This is not exactly what happens. I will talk through the
> > algorithm so
> > > that maybe we can find a good
> > > interface. I can probably write the code quickly:
> > >
> > > 1) We take in cells[numCells, numCorners], which is a list of all
> > the
> > > vertices in each cell
> > >
> > > The vertex numbers do not have to be a contiguous set. You
> can
> > > have any integers you want.
> > >
> > > 2) We create a sorted list of the unique vertex numbers on each
> > process.
> > > The new local vertex numbers
> > > are the locations in this list.
> > >
> >
> > Ok It took me re-writing this email a couple times but I think I
> > understand. I was too focused on local/global indices. But if I get
> > this
> > right, you still make an assumption: that the numVertices*dim
> > coordinates passed in vertexCoords are the coordinates of the
> > numvertices first vertices in the sorted list. Is that true ?
> >
> >
> > No. It can be arbitrary. That is why we make the vertexSF, so we can map
> > those coordinates back to the right processes.
> >
> So how do you know which coordinates correspond to which vertex since no
> map is explicitly provided ?
>
Ah, it is based on the input assumptions when everything is put together.
You could call BuildParallel()
with any old numbering of vertices. However, if you call CreateParallel(),
so that you are also passing
in an array for vertex coordinates. Now vertices are assumed to be numbered
[0, Nv) to correspond to
the input array. Now, that array of coordinates can be chopped up
differently in parallel than the vertices
for subdomains. The vertexSF can convert between the mappings.
> > > Here is my proposed interface. We preserve this list of unique
> > vertices,
> > > just as we preserve the vertexSF.
> > > Then after DMPlexCreateFromCellListParallelPetsc(), can
> > > DMPlexInterpolate(), you could call
> > >
> > > DMPlexBuildFaceLabelsFromCellList(dm, numFaces, faces,
> labelName,
> > > labelValues)
> > >
> > > Would that work for you? I think I could do that in a couple of
> > hours.
> > >
> >
> > So that function would be very helpful. But is it as simple as you're
> > thinking ? The sorted list gives local index -> unique identifier,
> but
> >
> > what we need is the other way round, isn't it ?
> >
> >
> > You get the other way using search.
> >
> Do you mean a linear search for each vertex ?
>
Since it is sorted, it is a log search.
Thanks,
Matt
> Thanks,
>
> --
> Nicolas
>
> > Thanks,
> >
> > Matt
> >
> > Once I understand better, I can have a first draft in my plexadapt
> code
> > and we pull it out later.
> >
> > Thanks
> >
> > --
> > Nicolas
> >
> >
> > > Thanks,
> > >
> > > Matt
> > >
> > > Is there a way to track this renumbering ? For owned
> > vertices, I can
> > > find the local index from the global one (so do old local
> > index ->
> > > global index -> new local index). For the remote ones, I'm
> > not sure. I
> > > can hash global indices, but is there a more idiomatic way ?
> > >
> > > Thanks,
> > >
> > > --
> > > Nicolas
> > >
> > >
> > >
> > > --
> > > 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/20210322/9b693627/attachment-0001.html>
More information about the petsc-users
mailing list