high performance computing on graphics processing units: hgpu.org

This paper describes an implementation strategy in preparation for an implementation of an OpenCL FFT. The two most essential factors (memory bandwidth and locality) that are crucial to obtain high performance on a GPU for an FFT implementation are highlighted. Theoretical upper bounds for performance in terms of the locality factor are derived. An implementation strategy is proposed that takes these factors into consideration so that the resulting implementation has the potential to achieve high performance.