Performance Portable pour le Calcul à Haut Débit

For the scientific community, and regardless of the field, the need for computing power and storage capacity is steadily growing. To meet this demand, multi-site computing infrastructures have been set up and equipped with a wide variety of hardware, including a growing number of GPU accelerators. Particle physics is no exception since it is increasingly diversifying its infrastructure with computing grids, supercomputers, as well as public and private clouds. For example, at CERN, the upcoming upgrade of the high-luminosity LHC, designed to better understand our universe, is expected to increase its computing needs by at least an order of magnitude over the next decade. More broadly, this means that science now progresses at the pace of available computing power. However, new advances in hardware and software often come at the expense of ease of programming. It is therefore becoming increasingly difficult to develop software with portable performance, which is necessary to fully exploit the hardware available at all these sites. The academic world and industry have therefore organized themselves around standards and proposed several software libraries. However, these libraries struggle to sustainably support the diversity and rapid evolution of hardware. In light of this observation, this thesis addresses cutting-edge techniques that abstract hardware complexity while maintaining the performance required for scientific applications. Here we present the new Kiwaku library, the outcome of our research, designed according to key principles of software engineering and advanced generic programming techniques, whose implementation is made possible by the recent C++20 standard. Kiwaku includes multidimensional containers and traversal algorithms that are of great relevance to the scientific community, easily adaptable to different execution contexts without compromise, as it offers portable performance comparable to that of case-by-case optimized code. We demonstrate this through a rigorous evaluation of Kiwaku's performance, both on synthetic benchmarks and on a particle physics use case.