Pipelined Asynchronous High Level Synthesis for General Programs

High-level synthesis (HLS) translates algorithms from software programming language into hardware. We use the dataflow HLS methodology to translate programs into asynchronous circuits by implementing programs using asynchronous dataflow elements as hardware building blocks. We extend the prior work in dataflow synthesis in the following aspects:i) we propose Fluid to synthesize pipelined dataflow circuits for real-world programs with complex control flows, which are not supported in the previous work; ii) we propose PipeLink to permit pipelined access to shared resources in the dataflow circuit. Dataflow circuit results in distributed control and an implicitly pipelined implementation. However, resource sharing in the presence of pipelining is challenging in this context due to the absence of a global scheduler. Traditional solutions to this problem impose restrictions on pipelining to guarantee mutually exclusive access to the shared resource, but PipeLink removes such restrictions and can generate pipelined asynchronous dataflow circuits for shared function calls, pipelined memory accesses and function pointers; iii) we apply several dataflow optimizations to improve the quality of the synthesized dataflow circuits; iv) we implement our system (Fluid + PipeLink) on the LLVM compiler framework, which allows us to take advantage of the optimization efforts from the compiler community; v) we compare our system with a widely-used academic HLS tool and two commercial HLS tools. Compared to commercial (academic) HLS tools, our system results in 12X (20X) reduction in energy, 1.29X (1.64X) improvement in throughput, 1.27X (1.61X) improvement in latency at a cost of 2.4X (1.61X) increase in the area