Exploring the Efficiency of the OpenCL Pipe Semantic on an FPGA

This paper evaluates the potential benefits of leveraging the OpenCL Pipe semantic to accelerate FPGA-based applications. Our work focuses on streaming applications in the embedded vision processing domain. These applications are well-suited for concurrent kernel execution support and inter-kernel communication enabled by using OpenCL pipes. We analyze the impact of multiple design factors and application optimizations to improve the performance offered by OpenCL Pipes. The design tradeoffs considered include: the execution granularity across kernels, the rate and volume of data transfers, and the Pipe size. For our case study application of vision ow, we observe a 2.8X increase in throughput for tuned pipelined kernels, as compared to non-pipelined execution. In addition, we propose a novel mechanism to efficiently capture the behavior for 2-dimensional (2D) vision algorithms to benefit Pipe-based execution.