Automatic parameter selection optimization based on the Triton compiler

The core of the autotune mechanism of the Triton compiler is to find the optimal solution from the solution space, that is, selecting the best parameter configuration; however, the process is currently overly reliant on manual tuning, which not only has poor performance but also significantly increases the compilation overhead when the solution space is too large. To address this problem, this study constructs a kernel-type parameter configuration algorithm and automates the parameter configuration selection. On this basis, a machine learning model is introduced to improve the model prediction accuracy and accelerate the acquisition of optimal solutions through feature selection and Grid Search. The experiments are tested on NVIDIA 3080Ti and A100 platforms, and after parameter tuning, the final model prediction accuracy can reach 85%, and the average acceleration ratio of model prediction in compilation time reaches 4.51. The experimental results show that the automation of parameter selection and model-guided selection can help reduce the compilation cost and achieve optimal kernel performance.