Optimizing the integration of charged particles on GPU

The upcoming HL-LHC runs are projected to require unprecedented demands on computing resources, with detector simulations representing a major part of the workload. To address these challenges, CERN is exploring GPU-accelerated particle transport through the AdePT project, a plugin to Geant4 which offloads the simulation of electrons, positrons and photons to GPU. In this report, we investigate possible performance optimizations for the charged particle integration. First, we experiment with reduced floating-point precision in the numerical integration loop. We show that single precision arithmetic provides significant speedups with negligible loss of accuracy in uniform magnetic fields. Second, we propose a new method for estimating distances to complex detector geometries, in particular for the electromagnetic calorimeter of ATLAS. We show that this estimator returns more accurate distance estimations which translates into an overall speedup of simulations.