Tobias, what you describe seems a bug in the library and we should try
to fix it. Maybe we can put it on the list for Heidelberg, though the
code jam will find me in the air over the Atlantic Ocean.
Every deal.II function (ab)using user pointers or indices should do the
following:
At beginning: figure out if indices, pointers or none have been used.
Dump the data if necessary.
At end: depending on the test at the beginning, restore or clear the data.
Alternatively, we might think if it is not better to enable users to add
arbitrary amounts of data and thus avoid the problem of dumping it.
Guido
Tobias Leicht wrote:
Hi Roman,
the problem is with the partitioning and does not depend on refinement
directly.
As you said, the user indices and pointers are supposed to be only for the
user. That's why any library functions store them first and restore them
afterwards if they use them internally. That worked fine as long as only user
pointers were available. Then user indices were introduced... Everything
would still work fine if we just copied all relevant data as it is, including
the information which type has been used lately. The actual implementation,
however, loops over all lines and quads and hexes (in 3D) of the mesh using
the accessor function to obtain the data and store it in a vector and set it
again the same way afterwards. Here, you have to decide whether you use
indices or pointers. The partitioning uses indices. If you used user indices
before, then everything is fine and unchanged afterwards. However, if you had
a clean triangulation with no indices and no pointers, after the partitioning
you have all zero indices and cannot use pointers without clearing all user
data first. On the other hand, if you had user pointers you cannot even
partition the mesh as it will fail on retrieving the user index of the first
line in the mesh.
We do not need a complicated example, by the way:
- take any mesh, set a user pointer and call partitioning -> fails
- take any mesh, partition and set a user pointer -> fails
Someone should implement a decent backup-and-restore functionality for general
user data, not indices or pointers.
Best, Tobias
_______________________________________________
dealii mailing list http://poisson.dealii.org/mailman/listinfo/dealii
