Sort ghost indices in sub-dofmap based on original ordering

[jorgensd]

Relates to: #2881, but now in the context of usage in #2929.
Currently, collapsing a dofmap of a blocked space states that is does not reorder the degrees of freedom.
However, the ghost dofs are re-ordered, as the new index map is created with create_sub_map.

This PR adds in a sorting of the ghost indices, based on the indices of the initial ghosts.
It also relates to #2919.

[francesco-ballarin]

So basically now we are moving to the position I used to have in #2881 (comment)?

I am fine either way, I am just asking because if we do guarantee that ghosts are now sorted as in the original ordering, I will beef up again my asserts 😛

[jorgensd]

So basically now we are moving to the position I used to have in #2881 (comment)?

I am fine either way, I am just asking because if we do guarantee that ghosts are now sorted as in the original ordering, I will beef up again my asserts 😛

Yes, it s either that or not use create_sub_map for collapsing spaces

[francesco-ballarin]

@jorgensd @jpdean need to decide which one you'd like to get merged first between #2890 and #2930, since they are changing the same functionality, and then update the other PR.

[garth-wells]

There are a few things to work through first, see #2929 (comment).

We want to aim for imposing as few constraints as reasonably possible, so ideally we don't want to force the re-ordering. If the caller needs a certain order then it should be possible for the caller to do the re-ordering.

[jorgensd]

There are a few things to work through first, see #2929 (comment).

We want to aim for imposing as few constraints as reasonably possible, so ideally we don't want to force the re-ordering. If the caller needs a certain order then it should be possible for the caller to do the re-ordering.

the points you have made in the referenced issue does not address the underlying issue, that if you create a sub index map of all indices, it ends up shuffling the ghosts.
The issue is that the shuffling of ghosts, when done on a collapsed map, completely changes the dofmap for the collapsed space of a blocked element, which does not make sense to me.
The re-ordering of a dofmap based on this index map is way more expensive than reordering prior to the «sub» index map construction.

[garth-wells]

Superseded by #3091