simple mesh-traversal procedure on parallel mesh

See https://answers.launchpad.net/dolfin/+question/218257 for explanation of the design.

Thanks Garth Wells, that solved my question.

decide_do_something = MeshFunction('bool', mesh, 0, False)

# set decide_do_something on each partition independently
for v in vertices(mesh):
    if all([cell_domains[c]==1 for c in cells(v)]):
         decide_do_something[v] = True

ngv = mesh.parallel_data().num_global_entities()[0]
g2lvi = mesh.parallel_data().global_to_local_entity_indices(0)

# sync values on shared vertices:
# if True on every partition we take True else False
for i in range(ngv):
    if g2lvi.has_key(i):
        temp = decide_do_something[g2lvi[i]]
    else:
        temp = True
    temp = MPI.min(temp) # little hack
    if g2lvi.has_key(i):
        decide_do_something[g2lvi[i]] = temp

# finally we can do something where True
for v in vertices(mesh):
    if decide_do_something[v] == True:
        do_something

Works in 1.0.0
Could be made much more efficient if we skip for-cycle for i which is not shared in no partitions..

1.0.0: ngv = mesh.parallel_data().num_global_entities()[0]
trunk: ngv = mesh.size_global(0)

But what about global_to_local_entity_indices map? Is it somehow accessible in trunk?

Thanks, Garth.

You can also build it this way:

l2gvi = mesh.topology().global_indices(0)
g2lvi = {}
for j in range(mesh.size(0)):
    g2lvi[l2gvi[j]] = j

Here l2gvi is local to global vertex index map (numpy.array) and g2lvi
is global to local vertex index map (dict).

Jan

Thanks Garth Wells, that solved my question.

On Sun, Jan 6, 2013 at 5:01 PM, Jan Blechta
<email address hidden> wrote:
> Question #218262 on DOLFIN changed:
> https://answers.launchpad.net/dolfin/+question/218262
>
> Status: Answered => Open
>
> Jan Blechta is still having a problem:
> Thanks, Garth.
>
> You can also build it this way:
>
> l2gvi = mesh.topology().global_indices(0)
> g2lvi = {}
> for j in range(mesh.size(0)):
> g2lvi[l2gvi[j]] = j
>
> Here l2gvi is local to global vertex index map (numpy.array) and g2lvi
> is global to local vertex index map (dict).
>

Take care with this approach because I'm working to de-couple the
local index from the array position (this is related to optimisations
for hyrbid MPI/ threaded operations). I don't know what will happen
with MeshTopology::global_indices(..).

Garth

> Jan
>
> --
> You received this question notification because you are a member of
> DOLFIN Team, which is an answer contact for DOLFIN.

> Take care with this approach because I'm working to de-couple the
> local index from the array position (this is related to optimisations
> for hyrbid MPI/ threaded operations). I don't know what will happen
> with MeshTopology::global_indices(..).

Ok. I assume that global_indices will stop working
instead of giving wrong values, won't they?

Regarding hybrid MPI/OpenMP do you consider creating some doc or demo
when it becomes useful? I'm interested in present state of usefulness
of approach as well as how to configure it. Do you know about any doc
or other source of information so far?

Jan