Author: Balasalle, J.
Paper Title Page
MOPPC089 CUDA Kernel Design for GPU-based Beam Dynamics Simulations 343
  • I.V. Pogorelov, K.M. Amyx, J. Balasalle, J. James
    Tech-X, Boulder, Colorado, USA
  • M. Borland, R. Soliday, Y. Wang
    ANL, Argonne, USA
  Funding: Work supported by the US DOE Office of Science, Office of Basic Energy Sciences under grant number DE-SC0004585.
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 CUDA kernels for transfer maps of single-particle-dynamics and collective-effects beamline elements, to be incorporated into a GPU-accelerated version of the ANL's accelerator code ELEGANT. In particular, we discuss techniques for efficient utilization of the device shared, cache, and local memory in the design of single-particle and collective-effects kernels. We also discuss the use of data-parallel and hardware-assisted approaches (segmented scan and atomic updates) for resolving memory contention issues at the charge deposition stage of algorithms for modeling collective effects. We present and discuss performance results for the CUDA kernels developed and optimized as part of this project.