Add support for mixed-dimensional kernels (codimension 1) #675
Conversation
There was a problem hiding this comment.
I think this PR looks good in general! I added a few remarks scattered across the code.
Another question is, could we ensure that this also works for co-dim 1 expressions? That shouldn't be alot of work, as the code-paths of integrands and expressions are the same.
| for domain in integrand.ufl_domains(): | ||
| if domain.topological_dimension() != cell.topological_dimension(): | ||
| is_mixed_dim = True | ||
|
|
There was a problem hiding this comment.
Is there a reason to use a boolean (is_mixed_dim) instead of computing codim = cell.topological_dimension() - domain.topological_dimension() ? Safer to check if codim > 0 rather than is_mixed_dim (which would be True even if codim < 0)
There was a problem hiding this comment.
is_mixed_dim indicates if an integral contains any mixed-dimensional terms. In build_optimized_tables, we need to differentiate between the codim 0 element and the codim 1 element in a mixed-dimensional integral so that only the quadrature rule for the codim 0 element is permuted (see here). I've added a check that codim > 0 when the codimension of the modified terminal is computed here. An alternative approach might be to replace is_mixed_dim with something like integral_codim and rename codim to element_codim. Let me know which you prefer
|
I think you've correctly noted that passing cell into For the analysis, I'd try to deduce the codim from cell and element and for the access you might need to store the codim as property of tabledata. Does that make sense or I am missing something? |
|
Thanks for the feedback everyone! This PR means that permutations are now only needed for the codim 0 cell, which simplifies this PR quite a lot. I think this will address most of the comments here, but I'll work on fixing the others soon. |
|
Are there any further issues anyone would like me to address? |
* Get working without perms * Tidy * Add TODOs * Add TODOs * Check if mixed dim * Permute tables * Ruff * Coeffs * Try with mt * Handle coeffs * Tidy * Ruff * Change to codim * Ruff * Add is_mixed_dim * Tidy * Don't permute facet element * Simplify * Tidy * Use domain * Add comment * Compute reference * Split data * Fix test * Docs * Docs * Remove TODO --------- Co-authored-by: Jørgen Schartum Dokken <dokken@simula.no>
This PR adds support for mixed-dimensional kernels of codimension 1. It requires this PR to work.
It would be nice to avoid passing
celltoanalyse_modified_terminal. Please let me know if anyone has any thoughts on this.