PC-Grade Parallel Processing and Hardware Acceleration for Large-Scale Data Analysis

Arguably, modern graphics processing units (GPU) are the first commodity, and desktop parallel processor. Although GPU programming was originated from the interactive rendering in graphical applications such as computer games,
researchers in the field of general purpose computation on GPU (GPGPU) are showing that the power, ubiquity and low cost of GPUs makes them an ideal alternative platform for high-performance computing. This has resulted in the
extensive exploration in using the GPU to accelerate general-purpose computations in many engineering and mathematical domains outside of
graphics. However, limited to the development complexity caused by the graphics-oriented concepts and development tools for GPU-programming, GPGPU has mainly been discussed in the academic domain so far and has not yet fully fulfilled its promises in the real world.

This thesis aims at exploiting GPGPU in the practical engineering domain and presented a novel contribution to GPGPU-driven linear time invariant (LTI) systems that are employed by the signal processing techniques in stylus-based or optical-based surface metrology and data processing. The core contributions that have been achieved in this project can be summarized as follow. Firstly, a thorough survey of the state-of-the-art of GPGPU applications and their
development approaches has been carried out in this thesis. In addition, the category of parallel architecture pattern that the GPGPU belongs to has been specified, which formed the foundation of the GPGPU programming framework
design in the thesis. Following this specification, a GPGPU programming framework is deduced as a general guideline to the various GPGPU programming models that are applied to a large diversity of algorithms in
scientific computing and engineering applications. Considering the evolution of GPU’s hardware architecture, the proposed frameworks cover through the transition of graphics-originated concepts for GPGPU programming based on
legacy GPUs and the abstraction of stream processing pattern represented by the compute unified device architecture (CUDA) in which GPU is considered as not only a graphics device but a streaming coprocessor of CPU. Secondly, the
proposed GPGPU programming framework are applied to the practical engineering applications, namely, the surface metrological data processing and image processing, to generate the programming models that aim to carry out parallel computing for the corresponding algorithms. The acceleration
performance of these models are evaluated in terms of the speed-up factor and the data accuracy, which enabled the generation of quantifiable benchmarks for evaluating consumer-grade parallel processors. It shows that the GPGPU
applications outperform the CPU solutions by up to 20 times without significant loss of data accuracy and any noticeable increase in source code complexity, which further validates the effectiveness of the proposed GPGPU general
programming framework. Thirdly, this thesis devised methods for carrying out result visualization directly on GPU by storing processed data in local GPU memory through making use of GPU’s rendering device features to achieve realtime interactions.

The algorithms employed in this thesis included various filtering techniques, discrete wavelet transform, and the fast Fourier Transform which cover the common operations implemented in most LTI systems in spatial and frequency domains. Considering the employed GPUs’ hardware designs, especially the structure of the rendering pipelines, and the characteristics of the algorithms, the series of proposed GPGPU programming models have proven its feasibility,
practicality, and robustness in real engineering applications. The developed GPGPU programming framework as well as the programming models are
anticipated to be adaptable for future consumer-level computing devices and other computational demanding applications. In addition, it is envisaged that the devised principles and methods in the framework design are likely to have
significant benefits outside the sphere of surface metrology.