Non-uniform DFT implementation for channel simulations in GPU

TLDR: A parallel scan based method to speed up channel simulation in wireless link-level OFDM network simulators without restricting the scope of the simulations is proposed, and the DFT properties in scan method are utilized to reduce register usage and hence the computation overhead of sine and cosine values.

Abstract: Channel simulation in wireless link-level OFDM network simulators involves a computationally intensive non-uniform discrete Fourier transform. In this paper, we propose a parallel scan based method to speed up this computation in GPU without restricting the scope of the simulations. We further utilize the DFT properties in scan method to reduce register usage and hence the computation overhead of sine and cosine values. This technique is compared against a method that saves computation by using uniform power delay profiles at the cost of generality, and we show that the performance is competitive. For single DFT, up to 19x speedup over a CPU implementation is observed using the scan based approach. For a simulation with 512 channels and a 1024 point DFT, the scan method gives a speedup of 141x with respect to the CPU, which compares favourably to the more restrictive uniform PDP method.