Next Generation Computational Fluid Dynamics (CFD) Codes

Accurate blast simulations based on first-principles, 3-D hydro-codes require hours on notebook PCs. Graphical processing units (GPUs) offer the potential of drastic reductions in run-times. Typical computational fluid dynamics codes perform of the order of 10-100 operations per memory fetch and store. In order to harness the potential of graphical processing units (GPUs), near-perfect streaming (or threading) of this information to and from processors is key to performance. During Phase I, we will develop optimal renumbering techniques for GPUs. These will then be incorporated into locfct.f, the core hydro-solver of FEFLO, a commonly used legacy code for blast-simulations. At the same time, locfct.f will be ported to the CPU via CUDA. Two typical blast simulation test cases will be computed and compared to the CPU implementation. The anticipated speedup is of the order of 1:8.