Author(s): Lennart Steffen; Reinhard Hinkelmann
Linked Author(s): Reinhard Hinkelmann
Keywords: Exascale; GPU computing; Parallel computing; Performance portability
Abstract: The newly developed C++ library Kokkidio allows its users to write expressive compute code with best-in-class performance across diverse hardware. By bridging the portability framework Kokkos with the math library Eigen, Kokkidio achieves both broad GPU support, as well as combined thread- and data-level parallelism on CPUs. An extensive evaluation shows Kokkidio to be up to 20× faster than Kokkos on CPUs, while being equivalent on GPUs. The comprehensive math functionality of Eigen allows hydroinformatics code written with Kokkidio code to be highly succinct and legible.
Year: 2026