Towards Portable Performance for Explicit Hydrodynamics Codes

Significantly increasing intra-node parallelism is widely recognised as being a key prerequisite for reaching exascale levels of computational performance. In future exascale systems it is likely that this performance improvement will be realised by increasing the parallelism available in traditional CPU devices and using massively-parallel hardware accelerators. The MPI programming model is starting to reach its scalability limit and is unable to take advantage of hardware accelerators; consequently, HPC centres (such as AWE) will have to decide how to develop their existing applications to best take advantage of future HPC system architectures. This work seeks to evaluate OpenCL as a candidate technology for implementing an alternative hybrid programming model, and whether it is able to deliver improved code portability whilst also maintaining or improving performance. On certain platforms the performance of our OpenCL implementation is within 4% of an optimised native version.