Pragma Based GPU Parallelizations for Cardiovascular Simulations

摘要

Directive based programming models as OpenACC, OpenMP 4.5 and CUDA provide a unified user interface for several hardware targets. We compare their performance on GPUs with respect to a solver for cardiovascular simulations using unstructured tetrahedral meshes and object oriented data structures in C++. Handling deep data copies appropriately in the constructors of the classes allows OpenACC to achieve the same performance as a well implemented CUDA solver. OpenMP 4.5 is still far away from reaching the same performance on GPUs. In contrast to OpenMP, OpenACC allows a user defined coalescing strategy and therefore OpenACC is already a realistic platform independent alternative to CUDA.