Optimizing Multithreaded Access to Global Variables in SPM through Compiler-Enhanced Dependency Analysis

초록

Scratchpad memory (SPM) occupies less space and consumes less energy than cache memory. However, unlike with cache memory, the user or the compiler must directly allocate the data or instructions. In this paper, we propose a new compiler based on the LLVM compiler that allocates frequently used global variables to the SPM and addresses issues in a multi-threaded environment with frequent I/O operations. The results demonstrate approximately a 27% reduction in energy consumption. © 2024 IEEE.