Transformations of High-Level Synthesis Codes for High-Performance Computing

Spatial computing architectures promise a major stride in performance and energy efficiency over the traditional load/store devices currently employed large scale systems. The adoption of high-level synthesis (HLS) from languages such as C++ OpenCL has greatly increased programmer productivity when designing for platforms. While this enabled wider audience to target spatial architectures, optimization principles known software design are no longer sufficient implement high-performance codes, due fundamentally distinct aspects hardware design, programming deep pipelines, distributed memory resources, scalable routing. To alleviate this, we present collection optimizing transformations HLS, targeting efficient (HPC) applications. We systematically identify classes (pipelining, scalability, memory), characteristics their effect on HLS code resulting (e.g., increasing data reuse or resource consumption), objectives that each transformation can resolve interface contention, increase parallelism). show how these be used efficiently exploit pipelining, on-chip fast memory, dataflow, allowing massively parallel architectures. quantify various transformations, cover process sample set HPC kernels, provided open source reference codes. aim establish common toolbox guide both engineers compiler tapping into potential offered by using HLS.