A Sampling-Based Approach to GPGPU Performance Auto-Tuning

摘要

We present a novel strategy for automatic performance tuning of GPU programs. The strategy combines heuristic search with regression trees, a machine learning model, to prune the optimization space. It samples configurations in the space as training data for a regression tree. It then focuses the search on the tree region with the best mean sample performance. Additional regression trees are built using the cumulative samples collected. This process is repeated until the given time budget is exhausted.We implement our strategy in OpenTuner, an open source automatic tuning framework. We evaluate the strategy using 8 benchmark GPU programs run on an Nvidia GTX 1060 GPU. We compare the effectiveness of our strategy in obtaining good performing configurations against the state-of-the-art AUC-Bandit strategy used by OpenTuner. Experimental results show that our strategy is more consistently able to obtain better performing configurations compared to OpenTuner's strategy, by up to 34% averaged over 100 auto-tuning experiments.