Recent changes to Forms?

[lindsayad]

I have a firedrake code that I haven't run in a few months. After updating firedrake I am getting the following stack trace:

Traceback (most recent call last):
  File "petsc4py/PETSc/libpetsc4py.pyx", line 1414, in petsc4py.PETSc.PCApply_Python
  File "/data/lindad/venv-firedrake/lib/python3.12/site-packages/firedrake/slate/static_condensation/sc_base.py", line 53, in apply
    self.forward_elimination(pc, x)
  File "/data/lindad/navierstokes-preconditioning/codes-hdg/navier-stokes/scpc2.py", line 328, in forward_elimination
    self._assemble_Srhs.assemble(tensor=self.condensed_rhs)
  File "/data/lindad/venv-firedrake/lib/python3.12/site-packages/firedrake/assemble.py", line 1001, in assemble
    self._check_tensor(tensor)
  File "/data/lindad/venv-firedrake/lib/python3.12/site-packages/firedrake/assemble.py", line 1234, in _check_tensor
    raise ValueError("Form's argument does not match provided result tensor")
ValueError: Form's argument does not match provided result tensor

Any ideas about how to fix/update our code? The scpc2.py that you see in the stack trace is our code which does static condensation for multiple fields (we have both facet velocity and pressure unknowns in our hybridized DG formulation). I looked in scpc.py which is the model for our PC for recent changes which brought me to #3947. I can't quite tell if that PR is the "culprit" or not. I'm happy to provide more details that might be helpful. I know as the question stands now it's not very detailed

[connorjward]

I think this issue was introduced some time ago when assemble was made more strict about whether it takes a Cofunction or Function. @JHopeCollins and @pbrubeck will know more.

[JHopeCollins]

This does look like it is due to the changes in #3947, specifically the changes from #3922 that were also merged in that PR.
The tensor argument to assemble now must be a Cofunction, whereas previously it was allowed to be a Function for backwards compatibility with pre-Cofunction code.

You can see that (what I assume is) the equivalent argument in firedrake.SCPC is now a Cofunction:

firedrake/firedrake/slate/static_condensation/scpc.py

Line 81 in 099ea9d

self.condensed_rhs = Cofunction(Vc.dual())

The default behaviour for dealing with boundary conditions was also changed in that PR - boundary conditions are assumed to already be enforced on the solution and the residual at boundary condition nodes is zeroed out - so you might have to make some changes there. You could compare your code to the the current firedrake.SCPC code to check if there's any differences in BC treatment.

On a separate note, if you have a version of SCPC that works for one or multiple fields then if you would be happy to contribute it we'd welcome a PR to add it to Firedrake.

[lindsayad]

Thanks for the reply @JHopeCollins! With your initial suggestion and combing more closely through scpc.py I was able to get our preconditioner working again. I do not think our approach to introduce a second field is very general; it would be best to allow an arbitrary number of fields. But yea I agree we should get a PR up at some point 👍

[JHopeCollins]

Great, good to hear you've got it working!