Monte Carlo simulation of particle collisions (event generation) will consume a significant fraction of computational resources at CERN’s High-Luminosity Large Hadron Collider. The CCE Event Generators (EG) team will develop, from scratch, a parallel matrix-element generator that runs on new and traditional architecture. This team will coordinate with the HEP Software Foundation EG group and efforts worldwide. In collaboration with PPS, it will explore multiple portable parallelization strategies for that.
Madgraph5 for GPUs: MadGraph5 is a commonly used event generator in LHC experiments, resulting in 10-20% of annual Worldwide LHC Grid usage. Being a smaller code library (of order 10k lines) than those of LHC experiments (of order 1M lines) and having focused computational kernels makes the task of reengineering for modern architectures easier to achieve. EG members are engaged with MadGraph developers to study the use of architecture independent frameworks like those studied in the PPS group, e.g. Kokkos, Sycl, and OpenMP.
Berends-Giele recursion on GPUs: Sherpa is a general-purpose event generator for collider physics, which provides much the same functionality as MadGraph + Pythia. The complete code base consists of about 150k lines, but most of the computing time is spent in kernels that number no more than 10k lines. These components are being redesigned in collaboration between EG members and Sherpa authors, as well as external collaborators. The project makes use of architecture independent frameworks such as Kokkos.