GUDA – A Lightweight Numerical Geometry Evaluation Library for Universal Device Accelerators

The widespread adoption of universal device accelerators, such as NVIDIA, AMD, and In- tel GPUs, in high-performance computing has revolutionized computational engineering and simulation, offering powerful platforms for executing complex numerical tasks. This paper presents GUDA (Geometry for Unified Device Acceleration), a software library designed to interrogate numerical geometry models directly on these device accelerators, eliminating the costly round-trip data migration between device and host typically required for evaluation. Built atop the open-source EGADSlite geometry kernel, GUDA provides essential func- tionality including geometry interrogation, inverse evaluation, and topology traversal. The library development prioritizes keeping geometric operations local to the device, minimiz- ing host-device communication overhead. This design enhances the efficiency of dependent engineering and simulation tools by leveraging on-device computation, offering a portable, high-performance solution. GUDA has been demonstrated on NVIDIA GPUs in a largely device-independent fashion. This paper explores GUDA’s architecture, implementation, and its impact on modern computational workflows.