On Mon, Feb 18, 2013 at 4:04 PM, Mark F. Adams <mark.adams at columbia.edu>wrote:
> Yes it should be removed, I just deleted the whole section and tested it. > > We could reinstate this code and check that all processors have a blocked > index set before making blocked ISs. > I don't want to add another MPI_Allreduce to the critical path (executed on each assembly). So we either make it strict enough to error if you declare a block size and don't fill in the blocks, we always fill in the blocks for you (maybe costing memory), we make the VecScatter permissive enough to have padded blocks, or we never optimize for the blocks. To answer this, we have to decide how much the block size attribute is about the _interface_ (MatSetValuesBlocked) and how much is about data structure optimizations. -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.mcs.anl.gov/pipermail/petsc-dev/attachments/20130218/3d2786b1/attachment.html>
