Python 'assemble' function is missing an argument for nonlinear problems

[Prof Garth Wells]

The Python assembler has the signature

def assemble(form,
             tensor=None,
             mesh=None,
             coefficients=None,
             function_spaces=None,
             cell_domains=None,
             exterior_facet_domains=None,
             interior_facet_domains=None,
             reset_sparsity=True,
             add_values=False,
             finalize_tensor=True,
             keep_diagonal=False,
             backend=None,
             form_compiler_parameters=None,
             bcs=None,
             symmetric_mod=None):

which is missing an argument to pass a vector x as is required inside a Newton loop.

Options are:

1. Add an extra argument, as with assemble_system; or
2. Remove the bc argument and let users apply the bcs (is there are compelling JIT-related issue why bcs should be applied inside assemble?

[Prof Garth Wells]

Marker is blocker - to be fixed before release.

[Johan Hake]

It looks like the bcs and symmetric_mode areleft over from the work of Joacim Haga wrt symmetric assemble. The commit message of 6c62604fed8b5274d7b5a2d1b7a683a2ffe54bb4 says that one can do:

A, An = symmetric_assemble(a, bcs=bc)
b = assemble(L, bcs=bc, symmetric_mod=An)
solve(A, x, b)

With symetric_assemble out of the code we should be able to just remove this.

It makes sense to have x0 in the assemble_system function but I am not sure it makes sense to introduce that to assemble as it only (more or less) mirrors the C++ version, which lacks functionality to pass an additional vector.

[Prof Garth Wells]

@johanhake Sounds good - let's remove bcs, and symmetric flagsfrom assemble, and add x0 to system_assembler.

[Johan Hake]

@garth.wells ok, will do. As this might potentially break some user code I suggest we add a deprecation message for the bcs arg.

Btw x0 is already present in assemble_system.

[Prof Garth Wells]

If we leave bcs, even as deprecated, should we add an x0 argument?

[Johan Hake]

There should not be any point to do that as the Assembler::assemble, which the python assemble use does not provide any interface for the x0 argument. bcs is now just used for convenience.

[Prof Garth Wells]

But if we let the Python assemble interface apply bcs, we should probably pass in x0 too.

[Johan Hake]

Not sure how that could be done? Assembler::assemble does (naturally) not support x0 and that is what the python assemble use. When bcs we introduced it was used to assemble symmetric rhs together with the symmetric_mod argument, which was used to pass a Matrix (not a flag), see syntax above.

Now bcs can only be used as convenience. In my view there are no compelling reasons why that should be provided, as it diverge the C++ and Python interface. But some users might rely on it in their code so therefore I suggest we keep it with a deprecation message. There is no demo or test using it.

[Johan Hake]

Fixed in: b13d8b6d6945524fcee5ae1db689d2bc6b63ef27

[Prof Garth Wells]

Removing milestone: 1.3 (automated comment)