Reproduce “Using hardware performance counters to speed up autotuning convergence on GPUs” 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
Jiří Filipovič, Jana Hozzová, Amin Nezarat, Jaroslav Oľha, Filip Petrovič
The method predicts operation counts from code and launch settings, then uses target-GPU counters to steer weighted random search, not runtime predictions. On GEMM, a GTX 1070-trained model found an RTX 2080 configuration within 1.1× of exhaustive-search best with 15.29× fewer empirical tests than random search. This measures search steps, not time; portability is demonstrated on tested cases.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.