Introducing a few local preconditioners by sv2518 · Pull Request #2259 · firedrakeproject/firedrake

sv2518 · 2021-10-25T14:23:30Z

This PR enables us to do specify some (local) preconditioners to the (local) operators in the local solves, used in the HybridizationPC.

All local preconditioners are controlled with PETSc options. The corresponding prefix is _lmi_ which stands for local mixed inverse. There are options to control which preconditioner to use and if it is replacing the original operator (preonly) or not. All of this is locally matrix-explicit only, the locally matrix-free PR will be coming in later.

The two options of local preconditioners are
a) a Jacobi preconditioner and
b) a user-supplied operator. Note that b) was originally in a separate PR, see #2240.

The first commit d8a3e15 is only refactoring and also renaming some existing options to look more PETSc like. Then there follow some infrastructural additions to Slate and its translations into lower levels, which I needed for the local diagonal preconditioners.

The last commit 10bdca7 is a big one. It includes adding more infrastructure into the SchurComplementBuilder class to deal with the new preconditioners, including the handling of possible options which are explained in detail in the SchurComplementBuilder class description. Testing is part of this commit too (I test both getting the right solutions and handling the options correctly in one go)

sv2518 · 2021-10-26T14:46:24Z

Okay this should be ready for review now.

firedrake/slate/static_condensation/hybridization.py

tests/slate/test_slate_hybridization.py

firedrake/slate/static_condensation/hybridization.py

tests/slate/test_slate_hybridization.py

firedrake/slate/slate.py

firedrake/slate/slac/utils.py

firedrake/slate/slac/compiler.py

firedrake/slate/slate.py

firedrake/slate/static_condensation/hybridization.py

wence-

I think this looks nice, and had a bunch of comments that might be helpful

firedrake/slate/slate.py

firedrake/slate/static_condensation/hybridization.py

…hur complement.

…ents a tensor of the same shape as A, but only carries the diagonal of A.

…ugh TSFC. The diagonal is returned in vector-shape.

…ate Tensors with the diagonal option.

… matfree application of diagonal preconditioners.

…s and the codegen for Slate's Reciprocal.

sv2518 · 2021-11-02T13:53:56Z

Okay I only need to clean my last few commits and then this is good to go? Thanks for the review @wence- and @dham!

…thin the HyrbidizationPC. The local preconditioner options include either a) jacobi or b) a custom operator supplied by the user.

…ossible

sv2518 · 2021-11-03T14:08:38Z

Are there any other issues or can we merge? I am pressing on this because a merge would make my work of preparing another PR easier.

wence-

Looks good, go ahead!

sv2518 · 2021-11-03T21:36:06Z

Ok cool, thanks! This is 5 commits behind master by now, do I need to rebase before I merge?

wence- · 2021-11-03T21:57:37Z

Ok cool, thanks! This is 5 commits behind master by now, do I need to rebase before I merge?

No, since the merge is clean you can just hit the button.

sv2518 mentioned this pull request Oct 25, 2021

User-defined preconditioning of local solver in HybridisationPC #2240

Closed

sv2518 force-pushed the sv/local-preconditioners branch 2 times, most recently from 60a3722 to 4ab62c6 Compare October 26, 2021 14:40

sv2518 marked this pull request as ready for review October 26, 2021 14:44

sv2518 commented Oct 28, 2021

View reviewed changes

firedrake/slate/static_condensation/hybridization.py Show resolved Hide resolved

sv2518 commented Oct 29, 2021

View reviewed changes