Tile-based Code Generation for Efficiently Accessing to Scratchpad Memory
Jaewook Lee, Yoonah Paik, Seon Wook Kim · 2021 International Conference on Electronics, Information, and Communication (ICEIC) · 2021
Most accelerators have used scratchpad memory (SPM) as software-managed working storage. However, the large amount of data used in recent applications like deep learning is hard to fit in the limited storage. It causes a significant burden to a programmer who should efficiently software-manage the SPM. This paper introduces a tile-based code generation for efficiently managing and executing an SPM by modifying Polly's loop tiling in LLVM, i.e., a polyhedral optimization library. For the performance analysis, we analyze the SPM copy overhead depending on a tiling factor.