OpEvo: An Evolutionary Method for Tensor Operator Optimization

Training and inference efficiency of deep neural networks highly rely on the performance tensor operators hardware platforms. Manually optimizing has limitations in terms supporting new or Therefore, automatically device code configurations is getting increasingly attractive. However, current methods for operator optimization usually suffer from poor sample-efficiency due to combinatorial search space. In this work, we propose a novel evolutionary method, OpEvo, which efficiently explores spaces by introducing topology-aware mutation operation based q-random walk leverage topological structures over spaces. Our comprehensive experiment results show that compared with state-of-the-art(SOTA) OpEvo can find best configuration lowest variance least efforts number trials wall-clock time. All work available online.