Accelerating Gaussian beam tracing method with dynamic parallelism on graphics processing units

Abstract

This study presents an efficient implementation of the Gaussian beam tracing (GBT) method utilizing graphics processing units (GPUs) to overcome the performance limitations of traditional CPU-based acoustic simulations. The algorithm was implemented and optimized on an NVIDIA RTX A6000 GPU, significantly enhancing the Gaussian beam summation (GBS) performance. We addressed the challenge of irregular control flows inherent to GBT by leveraging CUDA's dynamic parallelism to effectively flatten and dispatch nested loops directly on the GPU. Additionally, a profiling-driven optimization workflow using NVIDIA Nsight Compute enabled targeted improvements, raising SM throughput from 22.27% to 33.32%, L1 cache throughput from 13.15% to 22.15%, and L2 cache throughput from 9.16% to 21.26%. Consequently, the GPU-accelerated GBS algorithm achieved up to an 817× speedup compared to the original single-threaded CPU implementation, while the full computational pipeline reached 112× acceleration in a city-environment scenario involving 16,384 rays. Furthermore, this study introduces innovative strategies for overcoming GPU memory limitations, enabling efficient processing of large-scale ray datasets beyond single-kernel constraints. Finally, we establish systematic performance evaluation methodologies critical for analyzing and tuning GPU-accelerated algorithms, laying a foundation for future enhancements and scalability improvements.