Jed Brown <[email protected]> writes: > Matthew Knepley <[email protected]> writes: > >> Actually I now think I was wrong. Let me try to go through the use case, >> and you tell >> me where I am wrong: >> >> 1) We start with a system with block size B >> >> 2) We split it into k parts, so that \sum_k b_k = B, and b_k is the >> blocksize of each >> component. Isn't this how FS for structured grids works? > > Are you mixing up the size of the split with the size of the pieces? > That code path is only used when all the scalar splits are activated, > otherwise.
It looks like we dropped this thread. I think it requires adding block size arguments to the VecStride functions, or (better, IMO), stripping PCFIELDSPLIT of VecStride (replacing it with some use of VecScatter). Does it make sense to add a VecScatter interface that is optimized for doing several scatters at once (like VecStride), or is the reference implementation enough for now? If we change the VecStride interface, then it'll have to wait for the 3.5 release.
pgpQzoFw6BjZq.pgp
Description: PGP signature
