We recently presented new techniques for machine-learning-based predictive auto-tuning of GPU code at the inaugural Innovative Parallel Computing (InPar) conference in San Jose, CA.
Check out the paper here.