High-Performance Computing for Block-Iterative Tomography Reconstructions
|Mads Friis Hansen|
|Abstract||Modern equipment for X-ray tomography produces large amounts of data, and it is necessary to develop efficient high-performance algorithms and software for treating such problems. In this master thesis such algorithms that can take advantage of GPU accelerators are developed and implemented.|
Specifically, central processing unit (CPU) and graphical processing unit (GPU) kernels are developed in C++, implementing the projection method introduced by Joseph, for use with large-scale tomographic reconstruction problems in an existing framework.
The implementation is compared to other projection methods both with regard to reconstruction quality and computation performance. This is specifically oriented towards block-sequential and block-parallel versions of the row-oriented Kaczmarz algorithm (also known as ART), that can use the CPU and/or GPU to compute the forward- and back-projections without explicitly forming and storing the so-called system matrix. Block-sequential and block-parallel versions of the reconstruction algorithm will be compared to highlight the specific advantages and disadvantages to the different approaches, and an implementation of the block-sequential method, proven to be superior for multicore computing, is tested and analysed for the best performance with domain decomposition.
The work focuses on implementation aspects, including issues of efficiency and portability. Background regarding the theoretical foundation of the algorithms is also studied. The software is tested on large-scale experimental data from DTU and has performance studies and comparison of the chosen projection methods conducted.
|Type||Master's thesis [Academic thesis]|
|Publisher||Technical University of Denmark, Department of Applied Mathematics and Computer Science|
|Address||Richard Petersens Plads, Building 324, DK-2800 Kgs. Lyngby, Denmark, email@example.com|
|Series||DTU Compute M.Sc.-2016|
|BibTeX data|| [bibtex]|
|IMM Group(s)||Scientific Computing|