Improving HLS with Shared Accelerators: A Retrospective

This paper is a retrospective paper about our previous ICCAD2020 paper, ReconfAST: An Early Stage identification tool to find Shared Accelerators (SAs). SAs are specialized hardware accelerators that execute very different software kernels but share the common hardware functions between them. Our early detection methodology identifies computationally similar and synthesize-able kernels that are used to build SAs. Ourmethodology, ReconfAST, transforms one of the compiler’s intermediate output, the Abstract Syntax Trees (AST)s, into a new clustered AST (CAST) representation that further removes unneeded nodes and uses a regular expression to match common node configurations. SAs can provide increased coverage if both data flow and control flow similarities between seemingly very differentworkloads are detected. We saw a maximum reduction of above 200% in DSP, 75% reduction for LUTs, and 40% reduction for FFs compared to the smallest Dedicated Accelerators (DAs) with the best speedup. In this paper, we briefly explain our tool and discuss some challenges we experienced in building a tool that designs accelerators independent of a specific language. One example is the nonintuitive results HLS generates based on optimized mapping heuristics. We suggest ways to improve HLS tools and utilize solutions from other communities to help designers design and evaluate their systems methodically.