Paving the way to exascale for scientific applications
Date:
I gave the opening keynote at the 2022 IEEE/ACM International Workshop on Hierarchical Parallelism for Exascale Computing, during SC22.
I discussed the limitations of current C++ standards for HPC, highlighting the progress with std::mdspan that is in route for inclusion in C++23. I showed how current C++ parallel algorithms underutilize modern GPUs like the AMD MI250X and advocated for hierarchical parallelism (similar to Kokkos) to expose more parallelism. Finally, I explored the gap between the C++ memory model and hardware reality, showcasing how Kokkos, SYCL, and libcu++ extend std::atomic_ref with memory scope concepts.