Particle accelerator simulation using GPUs in Accelerator Toolbox

In storage ring light sources, two important performance parameters are injection efficiency and lifetime. Accelerator particle tracking in digital twins and simulators is an approach to efficiently derive these quantities and guide the optimization of the physical accelerator machine. These calculations are resource-intensive processes that often require high-performance computing. GPU usage can be an effective strategy to achieve good performance at a reasonable cost. This paper presents the current state of GPU particle tracking code implemented in Accelerator Toolbox (AT) using OpenCL or CUDA. The GPU kernel code is dynamically generated according to the lattice definition, allowing for optimal performance and selection of the symplectic integrator. We present accuracy and performance comparisons obtained with GPUs versus those obtained with AT CPU code. Trade-offs necessary to achieve optimal results are also discussed. Additionally, we discuss GPU kernel synchronization techniques needed by collective effect elements, as well as the object-oriented code structure to facilitate the integration of other GPU APIs.