Leveraging MLIR for GPU-accelerated stencil computing

Abstract

Weather and climate models extensively use stencil computations to model physical phenomena. These models have grown in complexity over the past decades and comprise large parts of today’s HPC workloads. With new architectures switching to heterogeneous compute nodes, climate models must be compiled and optimized for accelerator hardware. This presentation illustrates how a small team leveraged the MLIR infrastructure to compile stencil code for GPU accelerators. We demonstrate how we took advantage of the multi-level capabilities of MLIR by representing stencils in a high-level intermediate representation and subsequently lowering this representation down to GPU code. We show that the resulting toolchain outperforms the state-of-the-art when compiling the dynamical core of the European weather and climate model.

More details in the following paper: https://dl.acm.org/doi/10.1145/3469030