Preparing dawn for Weather and Climate Models on Triangular Grids

Dawn is a compiler-toolchain for domain-specific languages (DSLs) in weather and climate. Dawn addresses the large diversity of modern computing architectures by exposing a high level, descriptive DSL. The tool chain performs domain specific optimizations and subsequently generates efficient code for a number of computing architectures. This allows the same code to be re-used for multiple platforms without having to write architecture specific code or compiler directives. The efficiency of dawn has previously been shown by outperforming the expert tuned dynamical core of the COSMO model, both in performance and code size. However, dawn is currently limited to regular grid stencils (e.g. Finite Differences). In our current work, dawn is extended to be able to handle triangular meshes, enabling new classes of algorithms to be accepted by the tool chain, like the Finite Element and Volume Methods. As a case study, parts of the ICON model are implemented in dawn and first performance numbers on NVIDIA GPUs are presented.