Building Bridges: Julia as an MLIR Frontend

Driven by increasing compute requirements for deep learning models, compiler developers have been looking for ways to target specialised hardware and heterogeneous systems more efficiently. The MLIR project has the goal to offer infrastructure that can be used to develop new compilers and represent code at different levels of abstractions. While MLIR excels at offering developers a way to write new IR and transformations, there is no easy way for end users to generate code in these IR. In this work, we explore using the Julia programming language as a high-level input language for generating MLIR code. Most importantly, we focus on extensibility, allowing package developers to implement bindings to MLIR dialects in an intuitive and easy-to-use manner. By building on the Julia programming language, and its expressive features such as multiple dispatch and its extensible compiler, we design and implement a framework to generate MLIR code. Additionally, we evaluate this framework in three case studies. Ranging from developing a small DSL for einsum expressions, to specifying transformations on MLIR code and programming kernels to be run on GPU.