high performance computing on graphics processing units: hgpu.org

B-Splines are a useful tool in signal processing, and are widely used in the analysis of two and three-dimensional images. B-Splines provide a continuous representation of the signal, image, or volume, which is useful for interpolation, resampling, noise removal, and differentiation – all important steps in many signal processing algorithms. These splines are defined entirely by an array of coefficients that is roughly the same size as the original signal and of values in the same order of magnitude, making storage and representation trivial. What is not trivial, however, is the quick calculation and processing of those coefficients, especially for very large data. As technology improves in fields such as medical imaging, algorithms that use B-Splines will need to process increasingly higher resolution images and voxel volumes. New implementations are needed to make use of modern parallel architectures to keep these algorithms practical. This thesis presents a library for performing many common B-Splines operations in CUDA, the parallel programming framework for NVIDIA GPUs, and analyzes the considerations necessary when implementing a large-scale parallel version of such a well-established sequential algorithm. This library is meant to be used both by C++ programs as well as algorithms implemented in MATLAB without requiring significant changes. Significant speedups are obtained using this library to perform various common B-Spline image processing operations (as much as 30x for some), and the scalability limitations of the GPU implementation are addressed.