ESiWACE3 Deliverable D2.1 - Report on the tool development for an automated reduction in precision

This deliverable summarises the work performed on the reduction of numerical precision in weather and climate applications, with a focus on the NEMO ocean model and the new finite volume dynamical core developed at ECMWF – the IFS-FVM. This work has been carried out by the Barcelona SupercomputingCentre (BSC) and the European Centre for Medium-Range Weather Forecasts (ECMWF) as part of Task 2.1 of ESiWACE3. Building on prior efforts by the project partners to introduce reduced numerical precision through automated tools, significant progress has been achieved. Specifically, this deliverable documents work to improve the efficient use of the AutoRPE library that is developed at BSC and mainly applied to NEMO. This Python tool enables reading the NEMO source code, applying the precision reduction to a predefined precision level, and optimising the precision level using a divide-and-conquer algorithm. Additionally, the deliverable documents significant progress in enabling reduced numerical precision in the backend of the GT4Py Domain Specific Language (DSL), which facilitates the portability of IFS-FVM to modern supercomputing hardware, including GPUs. By introducing half precision in the backend, the first results using native half precision on the hardware were successfully obtained.