Improving Performance Portability of the Procedurally Generated High Energy Physics Event Generator MadGraph Using SYCL

Event Generators are essential tools for simulating Standard Model particle interactions, representing the initial step in modeling proton-proton collisions in the Large Hadron Collider (LHC) at CERN. Traditionally relying on a few algorithms like MadGraph5_aMC@NLO, there is a pressing need for enhanced portability across diverse computing platforms. In response, MadGraph5_aMC@NLO was expanded to support Nvidia GPUs using CUDA. This paper explores further advancements by porting MadGraph to SYCL, aiming to improve portability and leverage the heterogeneous architectures of modern supercomputers. Physics figures-of-merit are used to compare performance on Nvidia GPUs, AMD GPUs, and Intel GPUs and CPUs. A comprehensive performance comparison of MadGraph on Nvidia, AMD, and Intel GPUs, and Intel CPUs is presented, using physics figures-of-merit. This includes benchmarks against CUDA (on GPU) and OpenMP (on CPU) implementations, offering a holistic view of performance across different platforms. Weak scaling results on the Polaris (Nvidia GPUs) and Aurora (Intel GPUs) systems demonstrate that SYCL not only ensures portability but also maintains competitive performance. The adoption of SYCL achieves this without the complexities of managing multiple code bases or extensive use of preprocessor directives, presenting a promising approach for future developments in high-energy physics simulations.