Regression for interpolation of a callable on cells subset

[francesco-ballarin]

My RBniCSx CI started failing a couple of days ago, and I tracked down the issue to a regression in the interpolation of a callable on cell subsets.

The culprit can be either #2632, #2634, or #2636. I am leaning towards the first, but I can't really spot what has caused the regression.
I added a couple of failing tests in francesco/interpolation-callable-subset, feel free to use that branch for the fix. In particular:

• python/test/unit/fem/test_interpolation.py::test_interpolate_subset: the cases with callable_ == True are failing, while callable_ == False (i.e., the former version of the test, where a UFL expression is used rather than a callable) passes
• python/test/unit/fem/test_interpolation.py::test_interpolate_callable_subset: the case bound == 0.5 fails, while bound == 1.5 (i.e, the cell subset is actually the whole domain) passes

cc @niravshah241 since this regression may affect his work on shape parametrization/mesh motion

[jorgensd]

One thing to note in your tests is that you should scatter_forward after calling interpolation on a subset (not that it is the root cause of the error).

[francesco-ballarin]

Thanks @jorgensd . That call is indeed not present at the moment on main, but as you say it is not the root cause of the error, at least in serial.

[jorgensd]

Issue is:

    auto x_dofs = x_dofmap.links(cells[c]);
    auto x_dofs = stdex::submdspan(x_dofmap, c, stdex::full_extent);

In flatten dofmap. See https://github.com/FEniCS/dolfinx/blob/main/cpp/dolfinx/fem/interpolate.h#LL79
Fix is using

    auto x_dofs = stdex::submdspan(x_dofmap, c, stdex::full_extent);

not currently at my computer, so cant make a pr

[garth-wells]

@jorgensd good catch.