Performance evaluation of Kvazaar HEVC intra encoder on Xeon Phi many-core processor

This paper analyzes parallel scalability and coding speed of our open-source Kvazaar HEVC intra encoder on Intel Xeon Phi 61-core coprocessor that supports up to four hardware threads per core. The evaluated parallelization schemes of Kvazaar are 1) Wavefront Parallel Processing (WPP); and 2) tiles, both accelerated with picture-level parallel processing. With WPP, the C implementation of Kvazaar high-quality preset achieves an average speedup of 1.3 and a bit rate gain of 0.7% over the respective implementation of x265. Using tiles makes Kvazaar 1.4 times faster than x265 but at a cost of 0.3% bit rate loss. When high-speed presets are used, the speedup of Kvazaar increases to 1.4 with WPP and to 1.9 with tiles. Moreover, the respective coding efficiency of Kvazaar rises to 11.2% and 10.3%. Kvazaar also scales almost linearly to the number of cores in the processor. Even if the peak coding speed of Kvazaar on Xeon Phi is lower than that on the Intel 8-core i7 processor, our parallel scalability results promise excellent speed for Kvazaar on massively parallel processors equipped with more powerful cores.