AI-Driven Performance Metaprogramming - Embedding Spaces for Program Analysis and Optimization

Explore the intersection of artificial intelligence and program performance optimization in this 45-minute talk from the Scalable Parallel Computing Lab at ETH Zurich. Delve into the concept of embedding spaces and their application in representing complex information across various domains. Examine how these embeddings can be leveraged to assess, analyze, and enhance program performance. Learn about the evolution of program representation techniques, starting from textual LLVM IR embeddings for GPU execution time prediction to more sophisticated graph-based representations. Discover the advantages of graph-based approaches in capturing crucial relationships like data dependencies and flows. Investigate DaCe's performance metaprogramming capabilities and its programmable graph-based IR. Gain insights into the use of graph neural networks (GNNs) for creating performance embeddings that capture general performance properties. Understand how these embeddings can be utilized in Performance Embeddings for Transfer Tuning to select optimization metaprograms for IR graph transformation.