The way I've written MatFwk so far, it operates on nonoverlapping dof blocks, but the idea is that we want to first scatter a Vec into an "unrolled" version, duplicating the overlap dofs.
On Tue, Mar 16, 2010 at 1:26 PM, Jed Brown <jed at 59a2.org> wrote: > On Tue, 16 Mar 2010 12:51:53 -0500, Barry Smith <bsmith at mcs.anl.gov> wrote: >> If you use ILU(0) and the inplace option then it only needs one copy >> of the blocks instead of the two. > > With block Jacobi, it would be possible to avoid copying out the > sub-blocks by assembling into a matrix with the blocks already defined > (I think that Dmitry's MatFwk will allow this). ?Similarly for > Neumann-Neumann with many blocks per process. > > It would also be possible to assemble matrices with overlapping > subdomains, and define MatMult suitably (dropping redundant dofs). ?I > don't know whether Dmitry has plans to support this. > > Jed >
