AccFFT is a new massively parallel FFT library for CPU/GPU architectures. FFT is a fundamental algorithm originally developed by Cooley-Tukey. Many different libraries have been developed using this simple yet briliant algorithm. In recent years, the need to perform distributed, parallel computational has created a need for a fast parallel FFT library that can scale to thousands of processors. AccFFT is specifically designed with the goal of achieving maximum performance and scalability for both CPUs and GPUs. It uses a series of novel algorithms to reduce communication costs inherent in distributed FFTs.
We would like to thank Judith Hill and Dhairya Malhotra, who helped us with scaling the library on Titan and contributed to the library source code. AccFFT is funded primarily by US National Institutes of Health, US Department of Energy, US National Science Foundation, and US Air Force Office for Scientific Research. The specific funding sources are as follows:
- US National Institutes of Health/10042242
- US Department of Energy/DE-SC0010518
- US Department of Energy/DE-SC0009286
- US National Science Foundation/CCF-1337393
- US Air Force Office for Scientific Research /FA9550-12-10484
We would like to thank Oak Ridge Leadership Computing Facility (ORNL/OLCF) and Texas Advanced Computing Center (TACC) who garaciously provided us with computing time on their leadership supercomputers.