A Machine Learning Approach to Automatic Creation of Architecture-Sensitive Performance Heuristics

Recent interest in machine-learning based methods has produced many sophisticated models for performance modeling and optimization. These models tend to be sensitive to parameters of the underlying architecture and hence yield the highest prediction accuracy when trained on the target platform. Training a classifier, however, is a fairly involved process and requires knowledge of statistics and machine learning that the end users of such models may not possess. This paper presents a new framework for automatically generating machine-learning based performance models. A tool-chain is developed that provides automated mechanisms for sample generation, dynamic feature extraction, feature selection, data labeling, validation and hyper parameter tuning. We describe the design and implementation of this system and demonstrate its efficacy by developing a learning heuristic for register allocation in GPU kernels. Results show that auto-generated models can predict register thresholds that lead to integer factor performance improvements over kernels produced by state-of-the-art optimizing compilers.