JHopeCollins
commented
2 years ago I think this is ready to merge, or at least I've done all the refactors I wanted to do for this branch.
Turns out that most of the paradiag class was just dealing with the distributed all-at-once system, so most of the old paradiag methods have ended up in AllAtOnceSystem:
- all of the logic about dealing with the distributed functions - setting/getting components, timesteps, querying indices etc.
- updating the all-at-once function and the time halos from a PETSc
Vec. - constructing and evaluating the all-at-once form.
- transferring boundary conditions from a single timestep function space to the all-at-once function space.
Where possible, these methods can also take as arguments all-at-once functions belonging to someone else as arguments, so they can be reused by the JacobianMatrix and DiagFFTPC.
Other major points for this PR:
- Any dealing with the AAOS has been taken out of
paradiag.__init__(), except actually setting up the PETSc objects and solvers, because this is particular to the paradiag flavour being used. JacobianMatrixhas been moved to the same file as theAllAtOnceSystem, and neither class knows anything about theparadiagclass.JacobianMatrixgets all the information it needs fromAllAtOnceSystem. They do know about thealphathough so that we can still do a quasi-Jacobian (or reimplement Picard iterations at some point).- Renaming of a few variables for readability, eg
M->time_partition, andrT->time_rank
Factoring out the all-at-once system from the paradiag class.
AllAtOnceSystem encapsulates all data and logic related to a time-series spread over an ensemble - this include both the distrubuted function space/functions and the all-at-once finite element form.