Publication
In a context where the high-energy physics community strives to enhance software to handle increased data throughput, detector simulation is evolving to take advantage of new performance opportunities. Given the intricacy of particle transport simulation, recent advancements, such as adapting to accelerator hardware, require a significant research and development effort. The feasibility of porting complex particle transport codes to GPUs has been already demonstrated by parallelizing the processing of tracks undergoing electromagnetic interactions. However, the GPU workflow presents distinct performance bottlenecks compared to the CPU workflow, owing to differences in parallelism models and hardware capabilities. Notably, the geometry component in this workflow has emerged as the primary bottleneck that needs addressing to enhance GPU-based simulations. The current CUDA-aware geometry is a 3D-solid modeler featuring substantial warp divergence, primarily due to the varying complexity of the supported 3D primitive solid shapes. We present the outcomes of a one-year endeavor to create a bounded surface model that can seamlessly map the navigation capabilities of the existing solid-based geometry, with a focus on improving GPU efficiency. Our implementation aims to simplify low-level algorithms to reduce warp divergence, while simultaneously eliminating other sources of GPU inefficiency, such as recursive and virtual function calls, and ensuring code portability.