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

TL;DR

This paper tackles the computational bottlenecks of Gaussian Beam Tracing in complex environments by re-engineering the method for CUDA on an RTX A6000. It combines flat and dynamic parallelism to address irregular workloads and memory constraints, introducing a chunking mechanism to handle large ray sets. The approach achieves substantial speedups, up to about 790× in city environments and 188× in open fields, while validating accuracy against analytical solutions. The results indicate strong potential for real-time environmental acoustics, architectural acoustics, and environmental noise analysis, with clear guidance on future optimizations for the RT component and memory-management strategies.

Abstract

This study presents a reconstruction of the Gaussian Beam Tracing solution using CUDA, with a particular focus on the utilisation of GPU acceleration as a means of overcoming the performance limitations of traditional CPU algorithms in complex acoustic simulations. The algorithm is implemented and optimised on the NVIDIA RTX A6000 GPU, resulting in a notable enhancement in the performance of the Gaussian Beam Summation (GBS) process. In particular, the GPU-accelerated GBS algorithm demonstrated a significant enhancement in performance, reaching up to 790 times faster in city enviroment and 188 times faster in open plane enviroment compared to the original CPU-based program. To address the challenges of acceleration, the study introduce innovative solutions for handling irregular loops and GPU memory limitations, ensuring the efficient processing of large quantities of rays beyond the GPU's single-process capacity. Furthermore, this work established performance evaluation strategies crucial for analysing and reconstructing similar algorithms. Additionally, the study explored future directions for further accelerating the algorithm, laying the groundwork for ongoing improvements.

Figures (16)

• Figure 1: Schematic of a Gaussian beam and the associated coordinates.
• Figure 2: The programming flowcharts of the RT and GBS processes. The proposed dynamic parallelism targets to accelerate looping all rays in each process.
• Figure 3: (a) Typical CPU architecture; (b) Typical GPU architecture; Green indicates the processor while blue represents the memory; (c) Flowchart of the GPU acceleration implementation.
• Figure 4: The multi-threaded programming model of flat parallelism on GPU. These threads run simultaneously.
• Figure 5: The multi-threaded programming model of dynamic parallelism on GPU.