Issaquah (Feb 2023) US 45-106 24.7.8.3 presentation (default_accessor index type)

[mhoemmen]

Link to NB comment: https://github.com/cplusplus/nbballot/issues/522

Christian Trott created a benchmark to test the effects of the proposed change.

• Link: https://github.com/kokkos/code-examples/tree/main/stdcpp-experiments/mdspan/default_accessor_with_index_type

• NB's proposed change did not significantly improve performance on three different architectures (NVIDIA A100, AMD MI210, and AMD EPYC).

• Changing index_type in extents from a 64-bit type to a 32-bit type did improve performance. That was the change introduced into C++23 by P2553.

• The change would require nontrivial wording changes, given preconditions on index_type in the layout mappings. What happens if the index_type in the accessor doesn't match the one in the mapping?

• Therefore, we recommend against the NB comment (don't change C++23).

[mhoemmen]

LEWG outcome: Don't change C++23.