Abstract

We introduce Shamrock, a performance-portable framework written in C++17, targeting CPU and GPUs from any vendors using the SYCL programming standard, designed for numerical astrophysics across a wide range of hardware, from laptops to Exascale systems. Astrophysical schemes often share a common structure: a combination of neighbor searching and the numerical scheme itself. Shamrock embraces such abstractions to provide a common framework for multiple hydrodynamical schemes, namely finite elements, finite volume (with adaptive mesh refinement), and Smoothed Particle Hydrodynamics. To achieve this, at its core, Shamrock features a highly optimized, parallel tree algorithm with negligible construction overhead. This tree structure is coupled with a domain decomposition strategy that enables near-linear weak scalability across multiple GPUs. Shamrock achieves 92% weak scaling efficiency on 1024 AMD M I250x GPUs in large-scale Smoothed Particle Hydrodynamics (SPH) simulations. This corresponds to processing billions of particles per second, with tens of millions of particles handled per GPU , allowing us to perform the first SPH simulations above the billion-particle mark for protoplanetary discs.