Evaluating the graphics processing unit for digital audio synthesis and the development of HyperModels

The extraordinary growth in computation in single processors for almost half a century is becoming increasingly difficult to maintain. Future computational growth is expected from parallel processors, as seen in the increasing number of tightly coupled processors inside the conventional modern heterogeneous system. The graphics processing unit (GPU) is a massively parallel processing unit that can be used to accelerate particular digital audio processes; However, digital audio developers are cautious of adopting the GPU into their designs to avoid any complications the GPU architecture may have. For example, linear systems simulated using finite-difference-based physical model synthesis is highly suited for the GPU, but developers will be reluctant to use it without a complete evaluation of the GPU for digital audio. Previously limited by computation, the audio landscape could see future advancement by providing a comprehensive evaluation of the GPU in digital audio and developing a framework for accelerating particular audio processes. This thesis is separated into two parts; Part one evaluates the utility of the GPU as a hardware accelerator for digital audio processing using bespoke performance benchmarking suites. The results suggest that the GPU is appropriate under particular conditions; For example, the sample buffer size dispatched to the GPU must be within 32 to 512 to meet real-time digital audio requirements. However, despite some constraints, the GPU could support linear finite-difference-based physical models with 4X higher resolution than the equivalent CPU version. These results suggest that the GPU is superior to the CPU for high-resolution physical models. Therefore, the second part of this thesis presents the design of the novel HyperModels framework to facilitate the development of real-time linear physical models for interaction and performance. HyperModels uses vector graphics to describe a model's geometry and a domain-specific language (DSL) to define the physics equations that operate in the physical model. An implementation of the HyperModels framework is then objectively evaluated by comparing the performance with manually written CPU and GPU equivalent versions. The automatically generated GPU programs from HyperModels were shown to outperform the CPU versions for resolutions 64x64 and above whilst maintaining similar performance to the manually written GPU versions. To conclude part 2, the expressibility and usability of HyperModels is demonstrated by presenting two instruments built using the framewor