Efficient software-based runtime binary translation for coarse-grained reconfigurable architectures

Abstract

The increasing use of runtime-compiled applications provides an opportunity for coarse-grained reconfigurable architecture (CGRA) accelerators to be used in a user-transparent way. The challenge is to provide efficient runtime translation for CGRAs. Despite the apparent difficulties stemming from the heterogeneous nature of CGRAs, this paper demonstrates that it is possible to speed up runtime-compiled applications using CGRAs in a runtime-complied way. In particular this paper presents a runtime translation framework, called RBTVM, for CGRA accelerators, based on the LLVM Just-In-Time (JIT) compiler. Also two optimizations for the RBTVM are proposed. Experimental results show that the proposed RBTVM approach can improve the performance of runtime-compiled applications by 1.44 times on average compared to using the baseline JIT compiler only that does not take advantage of the accelerator, demonstrating the efficacy of the proposed approach.