Reproduce “Learning Generalizable Program and Architecture Representations for Performance Modeling” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Lingda Li, Thomas Flynn, Adolfy Hoisie
Researchers at Brookhaven National Laboratory developed PerfVec, a deep learning framework for performance modeling that learns distinct, generalizable representations for programs and microarchitectures. This approach achieves accurate performance predictions for unseen programs and hardware configurations, offering up to a 50x speedup over traditional simulation methods in design space exploration.
Key insights
- The execution time of a program can be accurately predicted by modeling its performance as a linear composition of individual instruction latencies and microarchitecture characteristics.
- Learning generalizable, microarchitecture-independent representations for instructions enables the creation of a "foundation model" that can predict performance for any program without retraining.
- Careful selection of microarchitecture-independent features, like stack distance and branch entropy, is essential for enabling a performance model to generalize effectively to unseen hardware.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.