GALA - A High Performance Graph Neural Network Acceleration Language and Compiler

Explore a 15-minute conference presentation introducing a domain-specific language and compiler designed to accelerate Graph Neural Network workloads through novel optimization strategies. Learn how researchers from the University of Illinois at Urbana-Champaign and University of Texas at Austin developed a system that addresses limitations in existing GNN frameworks by composing optimizations at both intra-operator and inter-operator levels. Discover the innovative approach that introduces two novel intermediate representations to track and compose transformations while maintaining a global view of GNN programs, including their training processes. Examine how this system enables training-specific transformations and achieves significant performance improvements, with geometric mean speedups of 2.55× for inference and 2.52× for training across multiple systems, graphs, and GNN models. Understand the domain-specific language that exposes intra-operator transformations as scheduling commands and the compiler's novel inter-operator transformations that exploit previously missed optimization opportunities in Graph Neural Network acceleration.