Porting a 2D Unstructured CFD code from the CPU to the GPU

Key words: graphics processing units, unstructured finite volume method, computational fluid dynamics. Traditionally, scientific computations are performed on Central Processing Units (CPUs) either in sequential or parallelized computing strategies. However, the recent appearance of powerful and massively parallel Graphics Processing Units (GPUs) is decisively contributing to its adoption in highly intensive scientific computations with concomitant reductions of computation times. It is known that GPUs are not as versatile as CPUs, but the recent availability of accessible programming tools and of GPUs able to operate in double precision create an opportunity for a leap forward in massively parallel computations [1]. Even though the most suitable applications for GPUs are those that feature high data parallelism with regular patterns to memory accesses, irregular patterns are not uncommon. Thus, this work explores the capabilities of GPUs to accelerate an unstructured 2D cell-centered finite volume computational fluid dynamics (CFD) code, which features irregular patterns arising from the unstructured grid. The linear momentum (1) and the continuity (2) equations are solved numerically for Newtonia