Study of Convolution Algorithms using CPU and Graphics Hardware

Abstract

In this thesis we evaluate different two-dimensional image convolution algorithms using Fast Fourier Transform (FFT) libraries on the CPU and on the graphics hardware, using Compute Unified Device Architecture (CUDA). The final product is used in VISSLA (VISualisation tool for Simulation of Light scattering and Aberrations), a software written in Matlab. VISSLATM is used to visualise the effects of cataracts, therefore it is important for our proposed method to be called from within Matlab. The product makes it possible to call graphics hardware from Matlab using the Mex interface. In this thesis we also explore the optimal usage of memory, and attempt to control allocated memory in a predictable way, to be able to minimise memory-related errors. A novel (hybrid) GPU/CPU algorithm using gpuArray and the row-column method is also presented and examined. Our proposed method speeds up the current computation on VISSLATM by 3-4 times. Additional proposed optimisations are examined along with the estimated resulting speedup.