Detail publikace
Accelerating Two-dimensional Wave Propagation Simulations Using Pruned Fast Fourier Transform
Spectral methods are highly efficient for solving wave propagation problems due to their use of Fourier transforms, which provide exponential convergence with respect to grid resolution, resulting in highly accurate gradient estimation. However, due to the complexity of material boundaries, oversampling of the medium is often required to prevent stair-casing artefacts and phase shifts. In cases where a narrow bandwidth source or weakly heterogeneous media are used-such as organs in the human body, rock formations in the earth's crust, or metals in engineering structures-the spatial spectrum pressure distribution tends to be sparse. Based on this observation, we hypothesise that the use of Pruned FFT can significantly improve computational efficiency. As a proof of concept, we implemented the pruned FFT for gradient calculation in ultrasound wave propagation benchmarks using the k-Wave toolbox. Our results, tested on human head models with varying materials such as skull, skin, and brain tissue, demonstrated a 1.8x speedup compared to the dense FFT version, with only a marginal increase in error of around 4% on average. This performance improvement shows promise for enhancing the efficiency of spectral methods in ultrasound simulations and other wave-based applications.
@inproceedings{BUT191375,
author="Ondřej {Olšák} and David {Bayer} and Jiří {Jaroš}",
title="Accelerating Two-dimensional Wave Propagation Simulations Using Pruned Fast Fourier Transform",
booktitle="2025 IEEE International Parallel and Distributed Processing Symposium (IPDPS)",
year="2025"
}