Comparison of parallel implementation strategies for the image source method for real-time virtual acoustics

The image source method (ISM) is inherently difficult to parallelize efficiently due to the data dependency of the image source tree. This paper presents and compares two different high-performance computing implementation strategies for the image source method. One strategy is designed such that the number of operations needed and the memory usage are reduced as much as possible using validity checks. However, this implies a data dependency that makes efficient parallelization impossible when scaling to massively parallel computing architectures, e.g., graphics processing units (GPUs). Thus, this method is mainly suited for multi-core CPUs, where the core count is low. The other strategy is designed for massively parallel computing architectures by eliminating data dependencies and conditional flow statements. However, in order to preserve data independence, considerable amount of unnecessary calculations must be carried out, some of which are ultimately discarded due to the image sources being invalid or invisible. The performance of the implementations is assessed when running the former method on a multi-core CPU and the latter on a GPU. Both implementations are found to be able to simulate within real-time constraints in fairly complex rooms up to a third reflection order, however, in most cases the GPU calculation is faster.