Chris Richardson

Seems fine. We still would want to be able to create a Vector just from an IndexMap too.

See branch chris/check-set-coeffs

This will require `MatrixCSR` to recompute its sparsity and some of its internal data (insert position for finalize for example). We also should check (in debug mode) that insertion to...

I've also used `pyamg` which is good.

https://github.com/FEniCS/dolfinx/tree/chris/petsc-optional is part way there. It is quite easy for C++ but will require more work for Python. Maybe we should proceed with the C++ part first?

I agree there is some discrepancy in the interface at the moment. If you want to experiment, I guess you could change the interface for `register_tabulate_tensor` to take a `std::function`....

:+1: - try to make something simple... it's not certain that this is the way to go. We may just use raw pointers throughout instead. But fine to experiment.

No. I am also not convinced it is always the right solution. Some solvers use a unified format, some split. I have got a working prototype (split format) so maybe...

https://github.com/FEniCS/dolfinx/tree/chris/matrixcsr-bs-experiment

I agree. I think we should consider supporting both formats internally.