Batched 3D-Distributed FFT Kernels Towards Practical DNS Codes
Toshiyuki Imamura, Aoki Masaaki, Mitsuo Yokokawa · Advances in parallel computing · 2020
This work introduces a new idea of batched 3D-FFT with a survey of data decomposition methods and a review of the states-of-arts high performance parallel FFT libraries.Besides, it is argued that the particular usage of multiple FFTs has been associated with the batched execution.The batched 3D-FFT kernel, which is performed on the K computer, shows 45.9% speedup when N and P are 2048 3 and 128, respectively.The batched FFT allows the developer to take advantage of a flexible internal data layout and scheduling to improve the total performance.