摘要
DNN compilers play a crucial role in mapping the loop nest onto accelerators. While existing approaches efficiently optimize these loop nests through multilevel optimization coupled with a convex solver, some case studies reveal a low robustness in handling data reuse across multiple NoC-interconnected engines. This issue is caused from hardware constraint violations during multilevel optimization. To address this challenge and enhance the robustness of the optimizer, this paper introduces a constraint tightening technique. The proposed technique achieves 100% solvability on end-to-end DNN optimization. With high solvability, especially on fused DNN layers, constraint tightening demonstrates a 1.14x EDP improvement.