| dc.description.abstracteng | With the dawn of the high luminosity era at the LHC, an unprecedented amount of data will be
collected and processed at the ATLAS experiment. This will result in a drastic increase of hit
combinatorics during track reconstruction with the ATLAS ITk detector, which will replace the
current Inner Detector. New algorithms and methods are investigated to efficiently process the
incoming data. One possibility is the deployment of hardware accelerators that provide massive
parallelism, like general-purpose computing on GPUs.
In this thesis, the detray library will be presented, for which a GPU-friendly tracking geom-
etry and navigation was developed. Within the ACTS (A Common Tracking Software) project,
which is a detector-agnostic toolkit of tracking algorithms written in modern C++, a dedicated
R&D effort was launched to investigate the adaptation of the ACTS tracking chain to GPUs.
The final GPU tracking demonstrator will provide a realistic setup of all steps of track recon-
struction, from clusterisation to ambiguity resolution, and thus allow an in-depth study of both
physics and compute performance of GPU-based tracking within the ATLAS experiment.
A crucial ingredient to be able to run track reconstruction is to ensure accurate and efficient
modelling of the detector geometry and its material. This is done in ACTS by the tracking
geometry, which is a purely surface based representation of the detector with a dedicated material
mapping step. The current implementation has been found in previous studies to have several
shortcomings concerning an adaptation to GPU computing, like its use of virtual function calls
to describe different geometrical shapes for the detector surfaces, or its use of vector-of-vector
data containers that rely on dynamic memory allocations.
With the detray library, a tracking geometry will be made available that solves these problems
by using a combination of static polymorphism and an index-based data management on global,
flat data containers in memory. Using the vecmem library for data management, the detray
detector can be read in from data files exported from existing tracking geometries in ACTS
on the CPU and subsequently be copied to the GPU memory system to run device-side track
reconstruction. The geometry description can be provided in full detail compared to ACTS,
including the access to material maps.
The detray tracking geometry and the track parameter navigation have been validated in
a constant magnetic field against a numeric approach using the Newton-Raphson algorithm,
enhanced with bisection steps, on several detector geometries, among them the current ITk
tracking geometry. | de |