Parallelising Wavefront Applications on General-Purpose GPU Devices

Pipelined wavefront applications form a large portion of the high performance scientific computing workloads at supercomputing centres such as LANL in the United States and AWE in the United Kingdom. This paper investigates the viability of utilising graphics processing units (GPUs) for the acceleration of these codes, using NVIDIA’s Compute Unified Device Architecture (CUDA). Wavefront applications differ from the massively data-parallel codes typically selected for execution on GPUs in that their computation must obey a strict data dependency, limiting the achievable level of parallelism. In this work, we identify a number of optimisations suitable for wavefront codes ported to this new architecture and attempt to quantify the characteristics of those codes that are most likely to experience speedups. Keywords-CUDA; GPU Computing; Wavefront; Hyperplane