Layout is part of the type: submdspan in numerical kernels on the CPU and the GPU
submdspan (C++26) lets a kernel work without index arithmetic on part of an array, such as a tile, a line, a plane or every other point. This report tries it on a tiled matrix product and transpose and on the kernels of three discretisations, chosen for their demanding memory patterns, on a CPU and a GPU. The views themselves cost almost nothing, and where they do cost something, the cost has a specific cause. What matters is how the data lies in memory: it changes performance by integer factors, often differently on the two devices, and the layout is usually fixed by changing a type rather than the kernel.
Authors
- Dan Vassilev
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-19
- DOI
- https://doi.org/10.5281/zenodo.24378315
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00