IMPAIR: Massively parallel regularised Richardson-Lucy deconvolution on heterogeneous hardware

This thesis investigates the comparative performance of multicore CPU and general purpose GPU on a commodity desktop computer. To investigate this, an image deconvolution software package (IMPAIR) was updated from its original cluster-computing design to support both of these parallel architectures. The IMPAIR software was chosen for this investigation due to the high memory and computational demands of the image restoration algorithms it implements, coupled with these algorithms’ natural amenity to highly parallelised solutions. IMPAIR performs the image deconvolution operation by parallelising either the unregularised Richardson Lucy algorithm (RL) or a wavelet regularised variant of Richardson Lucy (WRL), which carries a significantly higher computational cost but is more robust to the presence of high levels of noise in the algorithm’s input image. In order to support this WRL algorithm, general use wavelet shrinking libraries were developed for both the GPU and CPU, where a ×2 –×3 speedup of the GPU wavelet shrinking to the CPU wavelet shrinking was achieved. In total, eight parallelisation strategies for the IMPAIR deconvolution algorithms have been implemented and their runtime performance on a commodity desktop hardware is presented. Of the strategies presented, the “Topdown” multicore CPU strategy and the “Streaming” GPU strategy achieve similar runtimes, but the reduced memory footprint of the GPU Streaming strategy permits scaling up to image data over ten times the maximum capacity of the multicore CPU Topdown strategy, for both the regularised and unregularised Richardson Lucy algorithms