High Throughput 2D Spatial Image Filters on FPGAs (1710.05154v2)

Abstract: FPGAs are well established in the signal processing domain, where their fine-grained programmable nature allows the inherent parallelism in these applications to be exploited for enhanced performance. As architectures have evolved, FPGA vendors have added more heterogeneous resources to allow often-used functions to be implemented with higher performance, at lower power and using less area. DSP blocks, for example, have evolved from basic multipliers to support the multiply-accumulate operations that are the core of many signal processing tasks. While more features were added to DSP blocks, their structure and connectivity has been optimised primarily for one-dimensional signal processing. Basic operations in image processing are similar, but performed in a two-dimensional structure, and hence, many of the optimisations in newer DSP blocks are not exploited when mapping image processing algorithms to them. We present a detailed study of two-dimensional spatial filter implementation on FPGAs, showing how to maximise performance through exploitation of DSP block capabilities, while also presenting a lean border pixel management policy.