Framework for Transitioning DTRA High Fidelity Codes to APUs

The problem addressed by this proposal is the significant increase in runtimes needed to support DTRA high-fidelity simulations, even with availability of manycore CPU clusters. Simulation throughput is being bottlenecked by reliance on CPU-restricted algorithms with insufficient data parallelism exposed to take full advantage of modern Accelerated Processing Units (APUs). This effort seeks to develop a generalized performance analysis toolkit and template guidance methodology to assist developers in transitioning their codes to leverage APUs.  Framework development and assessment will emphasize software portability across a range of APU hardware vendors as well as provide facilitated guidance for use during code refactorization. A proxy flow solver, containing known bottlenecking code regions and memory access patterns associated with multi-physics simulations, will be used to assist in the evaluation and implementation testing of existing open-source APU programming frameworks. This includes meta-computing approaches such as the Kokkos library as well as directive-based computing approaches as offered by a variety of compiler vendors. Moreover, performance benchmarking of candidate frameworks will be conducted and evaluated using the capabilities of the Tuning and Analysis Utilities toolkit, which has been recently upgraded to better handle profiling diagnostics for software applications using APU parallelization frameworks, such as Kokkos.