High-level Synthesis Implementation of an Accurate HEVC Interpolation Filter on an FPGA

Abstract

This paper presents the first known high-level synthesis (HLS) implementation of an accurate interpolation filter for High Efficiency Video Coding (HEVC). The proposed multiplierless shift-register-based architecture is able to interpolate effectively up to four 8×8 blocks at a time for HEVC fractional motion estimation (FME). Our filter is implemented on Intel Arria V and Xilinx Virtex 6 FPGAs. On Arria V, it can operate at 270 MHz with 21.1 kALUTs. According to our profiling results, it can filter an adequate number of samples for FME in real-time 4K HEVC encoding of up to 85 frames per second (fps). On Virtex 6, the respective values are 313 MHz, 27.1 kLUTs, and 99 fps. The proposed solution doubles the speed over any of the existing interpolation filters for HEVC FME on an FPGA. It is also the only interpolation filter that meets the needs of real-time 4K HEVC encoder in practice and without any compromises in 23-bit filtering accuracy.