OpenCL Accelerated Multi-GPU Cone-Beam Reconstruction
Fraunhofer Institute for Production Systems and Design Technology IPK, Pascalstrasse 8-9, 10587 Berlin, Germany
The 12th International Meeting on Fully Three-Dimensional Image Reconstruction in Radiology and Nuclear Medicine, 2013
@article{kaseberg2013opencl,
title={OpenCL Accelerated Multi-GPU Cone-Beam Reconstruction},
author={K{"a}seberg, Marc and MelniN, Steffen and Keeve, Erwin},
year={2013}
}
Volume reconstruction in cone-beam CT is a computationally demanding task. Since recent years, the reconstruction is accelerated by utilizing Graphics Processing Units (GPUs). Frameworks for General Purpose Computations on GPUs are proven tool to access the resources of graphics cards. WIth the Open Computing Language (OpenCL) the first open standard for cross-vendor and cross-platform programming emerged, which allows to accelerate applications in heterogeneous environments. In this paper we present an implementation of an OpenCL accelerated volume reconstruction, based on the Feldkamp cone-beam CT algorithm. Our approach enables the utilization of multiple OpenCL devices in parallel. Furthermore the developed data management allows to handle volumes larger than the device memory. In experiments we proved the portability of our implementation on several devices from different vendors. Additionally, the performance scalability over multiple OpenCL devices was investigated. In a multi-GPU environment consisting of three NVIDIA GTX 580 our approach achieved up to 47.06 Giga Updates per Second and shows a speedup factor of 2.8 over a single GPU reconstruction.
December 17, 2013 by hgpu
