headerlogo
scyourway

SC Conference - Activity Details



Performance Programming Guidelines for Implementing FFT on GPU

Author:
Vasily Volkov  (University of California, Berkeley)
Posters Session
Tuesday,  05:15PM - 07:00PM
Room Rotunda Lobby
Abstract:
I present a prototype of the 1D complex-to-complex FFT on the GPU that outperforms any other GPU implementation and achieves up to 270 GFLOPS on a latest high-end GPU. It runs entirely in the on-chip memory for sizes up to N=1024 and is optimal on some of the recent GPUs, that is bound by peak memory throughput. This implementation is based on the same optimization guidelines as used in the current fastest implementation of the matrix-matrix multiply --- reducing data and thread parallelism to the necessary minimum, minimizing the use of the on-chip shared memory and using registers as the primary scratch space. I discuss performance analysis and modeling and how the performance scales across a number of GPUs with widely varying peak memory and arithmetic throughputs.
   IEEE Computer Society  /  ACM     2 0   Y E A R S   -   U N L E A S H I N G   T H E   P O W E R   O F   H P C