Heidelberg 2022 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 16: Experimental Methods (general) 1

T 16.7: Vortrag

Montag, 21. März 2022, 17:45–18:00, T-H29

detray - a GPU-friendly tracking geometry description — Andreas Salzburger¹, •Joana Niermann^1,2, Beomki Yeo^3,4, Attila Krasznahorkay¹, and Stan Lai² — ¹CERN — ²II. Physikalisches Institut, Georg-August-Universität Göttingen — ³Department of Physics, University of California — ⁴Lawrence Berkeley National Laboratory

With the next generation high luminosity experiments, the computational demand of particle track reconstruction will increase strongly. A potential way to tackle this is by offloading highly parallelizable tasks to an accelerator device. Existing codebases need to be adapted to e.g. specific host and device memory management and the calling of dedicated compute kernels, while avoiding code duplication as much as possible. Designed to be integrated into ACTS (A Common Tracking Software), which provides efficient algorithms for common tracking tasks, detray is an ongoing R&D effort to formulate the tracking geometry description for heterogeneous hardware. In order to propagate track states through the geometry model, detray follows the navigation design established in ACTS, but presents the geometry in a GPU-friendly way. It makes use of flat container structures without runtime polymorphism, a dedicated memory management scheme provided by the vecmem library, as well as direct indexing to link the geometry data, which together allows to instantiate the geometry model in host and device code. This talk gives an overview of the detray tracking geometry description and highlights the advantages and challenges of this GPU-friendly approach.