HyperModels - A Framework for GPU Accelerated Physical Modelling Sound Synthesis

Abstract

Physical modelling sound synthesis methods generate vast and intricate sound spaces that are navigated using meaningful parameters. Numerical based physical modelling synthesis methods provide authentic representations of the physics they model. Unfortunately, the application of these physical models are often limited because of their considerable computational requirements. In previous studies, the CPU has been shown to reliably support two-dimensional linear finite-difference models in real-time with resolutions up to 64x64. However, the near-ubiquitous parallel processing units known as GPUs have previously been used to process considerably larger resolutions, as high as 512x512 in real-time. GPU programming requires a low-level understanding of the architecture, which often imposes a barrier for entry for inexperienced practitioners. Therefore, this paper proposes HyperModels, a framework for automating the mapping of linear finite-difference based physical modelling synthesis into an optimised parallel form suitable for the GPU. An implementation of the design is then used to evaluate the objective performance of the framework by comparing the automated solution to manually developed equivalents. For the majority of the extensive performance profiling tests, the auto-generated programs were observed to perform only 6\% slower but in the worst-case scenario it was 50\% slower. The initial results suggests that, in most circumstances, the automation provided by the framework avoids the low-level expertise required to manually optimise the GPU, with only a small reduction in performance. However, there is still scope to improve the auto-generated optimisations. When comparing the performance of CPU to GPU equivalents, the parallel CPU version supports resolutions of up to 128x128 whilst the GPU continues to support higher resolutions up to 512x512. To conclude the paper, two instruments are developed using HyperModels based on established physical model designs.