Denis Demidov

Have you seen this? https://amgcl.readthedocs.io/en/latest/tutorial/poisson3DbMPI.html

I am sorry I do not understand you then. Shared-memory-parallel sounds like openmp to me: https://amgcl.readthedocs.io/en/latest/tutorial/poisson3Db.html

> instead it is a neutral/transparent interface/API to both SMP and MPI If you have MPI under the abstraction, maybe it is possible to somehow get the MPI communicator from...

You can just send local strips of the system matrix to each process (where each strip contains sequential chunk of rows, and the columns are globally numbered): https://github.com/ddemidov/amgcl/blob/720f6b7bc325a8a87df9ae33da125a80069b1d5c/amgcl/mpi/make_solver.hpp#L90-L99 The matrix...

> Does a row need to be entirely on a single proc? or could it be split across procs as well. Yes, each row has to be owned by a...

> is there a way to allow for this type of (block) partition in amgcl? No, there is not. > From what I reckon, internally in amgcl strips of rows...

> do I still need to ensure a contiguous chunk of rows belong entirely to a single processor? yes > Locating an entire row on a single processor appears to...

Sorry, I have not seen this error. Could this be a problem in CUDA SDK installation?

AMGCL may be used without boost, but most of the AMGCL examples depend on it, so it is unconditionally searched for in the CMakeLists.txt. It should be possible to make...

With 27ab1a7d4a6c7b0a2aa976c9f5b83bee0e06597c I can configure and use amgcl without Boost. Example of a client project that uses Boost-less amgcl: https://gist.github.com/ddemidov/74655be8a8a57aae861396c5a1a7ba40