JVET-X0104 Update on the progress of the optimized VVC encoder implementation, Ali266 [X. Dong, S. Fang, Z. Huang, J. Liu, S. Xu, R. Yang, L. Yu, J. Chen, R.-L. Liao, Y. Ye (Alibaba)]
This contribution reports on a software VVC encoder developed by Alibaba. Ali266 was first presented in JVET-W0127, it supports two presets, with its fast preset aimed at real-time encoding applications. In JVET-W0127, Ali266 was able to achieve encoding speed of 37.7 fps for 720p sequences.
This document reports further improvement in Ali266’s encoding speed at its fast preset, and also the addition of 8-bit encoding to Ali266 with the continued support of 10-bit encoding. It is reported that a 1.9x encoding speedup has been achieved in the past three months for video sequences at 1920x1080 or 1080x1920 resolutions. Currently, when configured with fast preset, Ali266 achieves average encoding speeds of 36.3 fps for classes 1920x1080 / 1080x1920 and 63.4 fps for classes 1280x720 / 720x1080. Based on such data, it was asserted that Ali266 is a VVC encoder capable of real-time encoding for 1080p sequences, and is more ready for real-world commercial applications than as described in reports on previous versions of this encoder.
Besides the speed improvement, Ali266’s encoding performance was also improved. For the JVET CTC content, Ali266 at slow preset achieves BD rates for {Y, U, V, YUV} of {10.95%, -6.11%, -4.50%, 7.11%} under the Random Access configuration with an average speedup factor of 227 times over VTM-13.0. For the internal VOD content, Ali266 at fast preset achieves BD rates for {Y, U, V, YUV} of {-36.25%, -45.73%, -43.02%, -38.10%} over the widely used open source HEVC software encoder x265 at medium preset, which is better than the result that was previously reported as {-32.87%, -46.07%, -42.70%, -35.35%} in JVET-W0127 over the same anchor.
How many threads were used? 12.
It was asked what are the typical bit rates for the fast preset where the company’s own content was used? This was not exactly known.
It was asked whether the fast preset results were report with 8 or 10 bit video? Both are possible, but the reported results were based on 8 bit.
The fast preset is still slower than X265 with its medium preset (about a factor of two).
Complexity analysis (0)
AHG10: Encoding algorithm optimization (3)
Contributions in this area were discussed in session 20 at 1300–1350 UTC on Wednesday 13 Oct. 2021 (chaired by JRO).