- Conference Article
- 10.1109/aiahpc66801.2025.11290366
Research on of Heterogeneous Parallel Algorithm for High-Dimensional FFT on GPGPU
- Sep 19, 2025
- Ruihan Sun + 3 more +3
The Fast Fourier Transform (FFT), as a fundamental algorithm in signal processing, has long been a focus of research on efficient computational methods. Given that General-Purpose Graphics Processing Units (GPGPUs) possess significantly more independent computing units than CPUs, they are well-suited for parallel computing tasks. Accordingly, an optimized radix-2 Stockham FFT algorithm based on the GPGPU architecture has been developed. The optimization employs a multi-step approach to enhance computational parallelism, incorporates a step-merging strategy to reduce kernel launch overhead and global memory access frequency, and utilizes the Stockham algorithm to improve global memory access patterns while minimizing shared memory bank conflicts. Experimental evaluation conducted on an Intel Xeon W5 and NVIDIA RTX A6000 platform demonstrated significant performance improvements over FFTW (v3.3.10). Maximum speedups of 5.40× for one-dimensional real-number transforms and 12.07× for multidimensional real-number transforms were achieved. For complex-number FFTs, speedups reached 6.52× (one-dimensional) and 19.06× (multidimensional).
Read more