Generic techniques in general purpose GPU programming with applications to ant colony and image processing algorithms

In 2006 NVIDIA introduced a new unified GPU architecture facilitating generalpurpose computation on the GPU. The following year NVIDIA introduced CUDA, a parallel programming architecture for developing general purpose applications for direct execution on the new unified GPU. CUDA exposes the GPU’s massively parallel architecture of the GPU so that parallel code can be written to execute much faster than its sequential counterpart. Although CUDA abstracts the underlying architecture, fully utilising and scheduling the GPU is non-trivial and has given rise to a new active area of research. Due to the inherent complexities pertaining to GPU development, in this thesis we explore and find efficient parallel mappings of existing and new parallel algorithms on the GPU using NVIDIA CUDA. We place particular emphasis on metaheuristics, image processing and designing reusable techniques and mappings that can be applied to other problems and domains. We begin by focusing on Ant Colony Optimisation (ACO), a nature inspired heuristic approach for solving optimisation problems. We present a versatile improved data-parallel approach for solving the Travelling Salesman Problem using ACO resulting in significant speedups. By extending our initial work, we show how existing mappings of ACO on the GPU are unable to compete against their sequential counterpart when common CPU optimisation strategies are employed and detail three distinct candidate set parallelisation strategies for execution on the GPU. By further extending our data-parallel approach we present the first implementation of an ACO-based edge detection algorithm on the GPU to reduce the execution time and improve the viability of ACO-based edge detection. We finish by presenting a new color edge detection technique using the volume of a pixel in the HSI color space along with a parallel GPU implementation that is able to withstand greater levels of noise than existing algorithms.