High-Level Synthesis of Pipelined FSM from Loop Nests

Embedded systems raise many challenges in power, space and speed efficiency. The current trend is to build heterogeneous systems on a chip with specialized processors and hardware accelerators. Generating an hardware accelerator from a computational kernel requires a deep reorganization of the code and the data. Typically, parallelism and memory bandwidth are met thanks to fine-grain loop transformations. Unfortunately, the resulting control automaton is often very complex and eventually bound the circuit frequency, which limits the benefits of the optimization. This is a major lock, which strongly limits the power of the code optimizations applicable by high-level synthesis tools.In this report, we propose an architecture of control automaton and an algorithm of high-level synthesis which translates efficiently the control required by fine-grain loop optimizations. Unlike the previous approaches, our control automaton can be pipelined at will, without any restriction. Hence, the frequency of the automaton can be as high as possible. Experimental results on FPGA confirms that our control circuit can reach a high frequency with a reasonable resource consumption