Instruction co-optimization based on triton compiler

Triton, as a high-performance compiler tailored for GPU computing, is capable of automatically converting high-level computations into efficient GPU code. Its optimization module includes classic optimization techniques such as instruction normalization, redundant instruction removal, and common subexpression elimination. However, there is still room for optimization regarding debugging instructions and specific computational instruction combinations. Addressing these issues, this paper proposes a collaborative optimization strategy targeting the intermediate representation of the Triton compiler, achieving performance enhancement through two key optimization techniques. Firstly, based on data flow analysis, it automatically identifies and safely removes redundant assert instructions generated during the model’s training or inference process. Secondly, through pattern matching, it converts low-efficiency instruction sequences into more efficient instruction combinations while strictly maintaining semantic consistency. Experimental results show that the average acceleration ratio of eliminating redundant assert instructions can reach 1.31, and the average acceleration ratio of instruction optimization can reach $\mathbf{1. 0 6}$. This collaborative optimization strategy has significantly improved the performance of code generated by the Triton compiler, providing more efficient execution efficiency for the training and inference of deep learning models.