DOI Bookmark: 10.1109/SCW72768.2026.00085

This paper describes the redesign of MDRangePolicy, the Kokkos Core abstraction for tightly nested multi-dimensional parallel loops, which improves performance on both GPU and CPU backends as of Kokkos 5.2.

Abstract:

This paper presents performance improvements for tightly nested multi-dimensional parallelism in the Kokkos library. We show how this can be implemented for GPU and CPU backends, which leads to improved performance and code maintainability. A widely used library like Kokkos needs to be updated in a manner that is friendly to the users and the developers. Using various benchmarks, we show how our work enables Kokkos users to write performance-portable code in a productive manner.

Recommended citation: Taberner et al. (2026). "Performance Improvement of Multi-Dimensional Parallel Loops in Kokkos." SC '26: Workshops of the International Conference for High Performance Computing, Networking, Storage and Analysis. Pages 378-392.
Download Paper