Preconditioned Richardson Iteration linear solver

[shreyas02]

Hi @JordiManyer, I was wondering why the Preconditioned Richardson iteration solver is not there and only a Richardon smoother is there . I have been using the Preconditioned Richardson method. Perhaps we could add this to the library by creating a pull request?

[JordiManyer]

Hi @shreyas02 this is covered by the LinearSolverFromSmoother solver. You can do

S = LinearSolverFromSmoother(smoother)

and it should work. It's not documented, nor tested. I want at some point to draw a clear distinction between smoothers and solvers, and have mappings between the two. This is a first step in that direction.

If you still think your PR brings something new, feel free to open it.

[shreyas02]

Hi @JordiManyer,
Thanks, I had a look at the code. It seems we can't pass arguments such as the tolerance to the solver. It always performs maximum iterations. For a solver, that is not ideal. Should I open a PR to add arguments such as tolerance and logging?

I understand that this does not make sense for a smoother, so I was thinking of making this argument optional and default setting to 0. So when used as a smoother, it performs n iterations; when used as a solver, it performs iterations until the required tolerance is achieved.

Also, could you briefly explain the purpose of LinearSolverFromSmoother(smoother)? I see that this is used just for performance benefits by preallocating memory.

[JordiManyer]

we can't pass arguments such as the tolerance to the solver

I see... I do not want any overhead in the smoothers if possible. Computing norms is relatively expensive in parallel. Considering this, I would prefer if you went ahead and implemented a RichardsonLinearSolver as proposed at the beginning.

Also, could you briefly explain the purpose of LinearSolverFromSmoother(smoother)? I see that this is used just for performance benefits by preallocating memory.

Its just to save lines of code. I just didn't want to maintain both the solver and the smoother. But I do understand that as a solver you probably want more control over the convergence. I overlooked that.

So yeah, go ahead and create a new LinearSolver. Tag me and I'll review the PR when it's done. Thank you very much!

[JordiManyer]

Closed by #87