PerfDojo - Automated ML Library Generation for Heterogeneous Architectures

Explore a research presentation introducing PerfLLM, a novel automatic optimization methodology that leverages Large Language Models and Reinforcement Learning to generate high-performance machine learning libraries for diverse hardware architectures. Learn about PerfDojo, an innovative environment that frames optimization as a reinforcement learning game using human-readable, mathematically-inspired code representations that guarantee semantic validity through transformations. Discover how this approach addresses the significant challenges of achieving optimal performance across the increasing complexity of machine learning models and the proliferation of diverse hardware architectures including CPUs, GPUs, and specialized accelerators. Understand how the methodology overcomes the limitations of manual optimization and existing automatic approaches that rely on complex hardware-specific heuristics and uninterpretable intermediate representations. Examine the system's ability to enable effective optimization without requiring prior hardware knowledge, facilitating both human analysis and RL agent training while achieving significant performance gains across diverse CPU architectures (x86, Arm, RISC-V) and GPU platforms. Gain insights into how this research addresses heterogeneity in instruction sets, specialized kernel requirements for different data types and model features such as sparsity and quantization, and architecture-specific optimizations that complicate performance tuning in modern computing environments.