search

SciML / LinearSolve.jl

LinearSolve.jl: High-Performance Unified Interface for Linear Solvers in Julia. Easily switch between factorization and Krylov methods, add preconditioners, and all in one interface.
https://docs.sciml.ai/LinearSolve/stable/
Other
249 stars 53 forks source link

AMGCL wrapper? #108

Open learning-chip opened 2 years ago

learning-chip commented 2 years ago

https://github.com/ddemidov/amgcl

Has OpenMP/MPI/CUDA implementations, similar to Hypre but much more lightweight. Already has a Python interface (see test)

ChrisRackauckas commented 2 years ago

Yes, that would make sense here.

j-fu commented 9 months ago

WIP: see https://github.com/JuliaPackaging/Yggdrasil/pull/7896

AMGCL is heavily templated. The best path forward would be a complete re-implementation in Julia which is over my time budget in the moment. That would be a great project for anyone wanting to learn precondtioning, though. In that case, the iterative solver part probably could be taken from Krylov.jl .

So for now any calls from julia need to go via a c library with pre-instantiated code, this is the above mentioned PR. Still quite bit flexibility though due to the "runtime interface". Makes scalar and point block variants available for OpenMP in the moment, but misses e.g. saddle point stuff and CUDA.

Will first register AMGCLWrap.jl, then make a PR over here.

j-fu commented 9 months ago

Ok, AMGCLWrap.jl is there.

It appears to be quite straightforward to produce an AMGMCLWrapLinearSolveExt via the LinearSolveFunction API.

Anything speaking for an LinearSolveAMGCLWrapExt instead ?

ChrisRackauckas commented 9 months ago

I don't care either way. However it's done, I think it needs a note in the LinearSolve.jl docs and with that the details of how the wrapper is done doesn't matter to the end user.

j-fu commented 9 months ago

Ok so will use the first version because it is easy to maintain for me. After getting some feeling how stable it works in some of my projects I will make a PR to the docs.