Multi-GPU

Support single system, multi-GPU configurations

• Based on OpenMP (for cross-platform support)
• Co-array model for inter-gpu communication
• Lightweight thread management, some basic multithreading routines (reductions, barriers, etc.)
• OpenCurrent philosophy: static communication patterns for efficiency. Solvers work unchanged (operate on local grids). Equations manage communication between multiple solvers.

MPI-enabled

Support for clusters

• Built on top of multi-gpu framework. New backend for the co-array model.

Templated operators

Allow for 'apply to grid'-type operations

• Built using C++ templates, and possibly metaprogramming (pending compiler support).
• Allow for more natural expression of per-cell operations, remove most of boiler-plate code.
• rewrite entire library using this framework - should shrink codebase considerably.

Solver enhancements

Implement better stencils

• Lots of good papers on efficient stencils in cuda. Implement some of them.

Iterative solvers

• New multigrid solvers for Poisson equations
• connect to cusp iterative solvers
• multi-gpu / multi-node support for all such solvers

Application areas

Ocean modeling

• Add sigma coordinate support to different solvers
• vertically integrated N.S. equations (2d)
• baroclinic / barotropic mode splitting solver

Others? (If interested, send mail to the list)