I read that on Mastodon, but it was the end today. I don’t know if videos or slides will be put online:
https://pasc24.pasc-conference.org/session/?sess=sess127
Description: Fortran, the primary programming language underpinning many operational weather and climate codes, was built around the fundamental principle that performance optimisation is left to the compiler. However, with the emergence of GPU accelerators, significant refactoring, often beyond simple addition of pragmas, is needed to achieve good GPU performance on established, operational, vectorised CPU code. This has led to the rise of DSLs and source-to-source methods that often use elements from compiler theory to bridge the CPU-GPU gap, leaving one unspoken question unanswered: Why the FORTRAN does my compiler not do this for me? In this minisymposium we aim to explore this question by looking at ECMWF’s CLOUDSC benchmark - an NWP mini-app designed to assess (and torture) compilers. For this benchmark, many GPU-optimised flavours exist, including Fortran and C-based offload flavours (OpenACC, OpenMP, CUDA, HIP), that provide an established performance baseline on different GPU architectures. Instead of further optimising these with more intrusive code changes, we ask the question “How close to the original vector-style Fortran code can we get without sacrificing performance?” We aim to explore this question with technologists and compiler enthusiasts from across the HPC and academic spectrum.