CUDA Kernel Design for GPU-Based Beam Dymanics Simulations
Tech-X Corporation, Boulder, CO 80303
International Particle Accelerator Conference (IPAC), 2012
@article{amyx2012cuda,
title={CUDA KERNEL DESIGN FOR GPU-BASED BEAM DYMANICS SIMULATIONS},
author={Amyx, K. and Balasalle, J. and Pogorelov, I. and Borland, M. and Soliday, R. and Wang, Y.},
year={2012}
}
Efficient implementation of general purpose particle tracking on GPUs can result in significant performance benefits to large scale particle tracking and tracking-based accelerator optimization simulations. We present our work on accelerating Argonne National Lab’s accelerator simulation code ELEGANT [1, 2] using CUDA-enabled GPUs [3]. In particular, we provide an overview of beamline elements ported to GPUs and briefly discuss optimization techniques for efficient utilization of the device memory space, with an emphasis on register usage. We also present a novel hardware-assisted technique for calculating a large histogram, and compare this to data-parallel implementations. In addition, we provide an overview of recent work on a new build system for ELEGANT and integrating GPU-accelerated elements with the existing codebase in a manner that allows ‘silent’ support of GPU acceleration. We conclude the paper with results of a realistic test simulation and comments on future work related to GPUenabled version of ELEGANT.
July 27, 2012 by hgpu
