WasserBayesTuner: A Synergistic Optimization Framework for Compiler Auto-Tuning via Wasserstein Metric and Bayesian Sampling

Addressing the dual challenges of efficient highdimensional parameter space exploration in modern compiler auto-tuning and the exploration-exploitation imbalance caused by traditional KL divergence’s sensitivity to zero-probability regions, this paper proposes WasserBayesTuner - a collaborative optimization framework integrating Wasserstein distance with Bayesian sampling. The methodology innovatively employs geometry-driven Wasserstein metrics to quantify configuration distribution discrepancies, effectively mitigating sensitivity to sparse probability regions, while establishing Bayesian Thompson Sampling mechanism for adaptive dynamic exploration strategy adjustment. Evaluations on the cBench benchmark demonstrate that the framework achieves an average speedup ratio of $1.37 \times$ under constrained tuning budgets, with $68.4 \%$ performance gains attained during critical optimization phases (within the first 40 minutes). Cross-task performance variation exhibits reduced standard deviation ($5.5 \%$), significantly outperforming conventional methods. Experimental results validate the framework’s robustness and convergence efficiency in highdimensional sparse optimization scenarios. Its generalized architecture establishes new directions for compiler parameter optimization, with future extensions to hyperparameter tuning and resource scheduling through meta-learning-driven crossplatform knowledge transfer, offering novel pathways for performance enhancement in heterogeneous computing ecosystems.