Introduction

Julia on GPUs for fun and profit

• Valentin Churavy

The conventional wisdom states that high-level dynamic programming languages are seldomly performant, ill-suited for high-performance computing, and impossible to execute on GPUs. If one wants to use a high-level language in HPC, it is merely for sugar-coating over some C++/Fortran behemoth or as an interface to machine learning libraries. Yet, one small village of indomitable optimists still holds out against the C++ hegemony. Julia is a high-level dynamic programming language created with scientific computing and performance in mind. It provides a GPU computing stack that compiles Julia directly to various GPU architectures. In this talk, I will cover the basics of GPU computing in Julia, how to write portable GPU kernels and programs, and give some insight into how the compiler stack behind it all works.