Hi, >> I want to use KSPGMRES
With what inside? Jacobi? Your outer solve was GMRES, which requires a linear preconditioner. That implies fully converging the subdomain solver when using an iterative method. This is likely to be inefficient, in which case you'd probably want to switch to FGMRES or GCR for the outer KSP. I don't know how difficult it will be to make BJacobi support GPUs. Could be easy, but someone has to write the code. Karl or Dominic, have you looked at this?
It depends on the amount of stability one wants to put it. If Gauss without pivoting is sufficient, Block-Jacobi shouldn't be too hard for the block size of interest (9-by-9). However, I suppose we do want robust solves, so it's better to use some explicit solution formulas. Cramer's rule is okay up to 4-by-4 blocks or maybe 5-by-5, but likely to be a mess for 9-by-9.
Best regards, Karli
