PHAST Library - Enabling single-source and high performance code for GPUs and multi-cores

The simulation of parallel heterogeneous architectures such as multi-cores and GPUs sets new challenges in the programming language/framework domain. Applications for simulators need to be expressed in a way that can be easily adapted for the specific architectures, effectively tuned for on each of them while preventing from introducing biases due to non-uniform hand-made optimizations. The most common heterogeneous programming frameworks are too low-level, so we propose PHAST, a high-level heterogeneous C++ library targetable on multi-cores and Nvidia GPUs. It permits to write code at a high level of abstraction, to reach good performance while allowing for fine parameter tuning and not shielding code from low-level optimizations. We evaluate PHAST in the case of DCT8x8 on both supported architectures. On multi-cores, we found that PHAST implementation is around ten times faster than OpenCL (AMD vendor) implementation, but up to about 4x slower than OpenCL (Intel vendor) one, which effectively leverages auto-vectorization. On Nvidia GPUs, PHAST code performs up to 55.14% better than CUDA SDK reference version