GPU acceleration of compton reconstruction for the PEDRO
School of Physics, Monash University, Melbourne, VIC 3800, Australia
IEEE Nuclear Science Symposium Conference Record (NSS/MIC), 2010
@inproceedings{dimmock2010gpu,
title={GPU acceleration of compton reconstruction for the PEDRO},
author={Dimmock, M.R. and Nikulin, D.A. and Brown, J. and Nguyen, C.V. and Gillam, J.E.},
booktitle={Nuclear Science Symposium Conference Record (NSS/MIC), 2010 IEEE},
pages={2665–2668},
year={2010},
organization={IEEE}
}
Compton reconstruction requires the computationally intensive, yet highly parallelizable, task of Cone of Response (CoR) back-projection. The acceleration of CoR back-projection is of significant importance as a faster algorithm allows the user to increase either the size or resolution of the imaging volume. Such acceleration also lends itself to the realization of real-time reconstruction. The image estimate is formed by calculating and histogramming voxel-cone intersections in the imaging volume. As the level of computation increases with the cube of the length or resolution of the volume, the bottle-neck in processing the dataset is at the imaging-voxel level, as opposed to the event level. The reconstruction algorithm developed for the Pixelated Emission Detector for RadiOisotopes (PEDRO) system was optimized for threading at the voxel level and implemented in the C++ programming language. The benchmark C++ code was then extended with OpenCL to assess the acceleration that could be obtained by running the algorithm on a selection of commercially available Graphical Processing Units (GPUs). Increases of up to 102 times the bench-mark speed were achieved
June 21, 2011 by hgpu
