Strategy with precs ?

[j-fu]

Question❓

What is should be the strategy with implementing precs ?
The pattern I like is:

struct ILUZero_ILU0 end

function (::ILUZero_ILU0)(A, p=nothing)
    (ilu0(sparse(A)),I)
end

so we can have a solver like KrylovJL_CG(precs=ILUZero_ILU0()) etc.
Where this implementation should live ?

a) Due to the fact that this implementation does not depend on LinearSolve, it could become part of the respectve preconditioning packages. So should we go and make PRs there ?

b) Or do you plan to have Package extensions for LinearSolve ?

I find a) quite intriguing...

[oscardssmith]

I think for now, it belongs in the preconditioner package, but they could also belong as package extensions. Both approaches seem to work to me, and I don't think either has a hidden downside.

[j-fu]

Just wanted to follow up here, now that the code has settled. I intend to do a couple of PRs to ILUZero, AlgebraicMultigrid and the like (i.e. those that I "like"). As for a consistent pattern, I would choose something like ILUZero.ILU0Precs, AlgebraicMultigrid.AlgebraicMultigridPrecs etc., i.e. introduce the pattern XXXPrecs, so people would write KrylovJL_CG(precs=ILU0Precs). Other naming like XXXPreconditioners might be an alternative at this point in time.