JVET-S0224 AHG16: Performance of a reasonably fast VVC software decoder [F. Bossen (Sharp)]
This contribution was discussed during 2010–2030 on 26 June (chaired by JRO)
Performance of a reasonably fast VVC software decoder is described: it is asserted to run about 30 times faster than the VTM on an 8-core processor when processing 4K bitstreams (CTC RA). No claim of optimality was made, as the optimized decoder was written from scratch in a short amount of time.
The table below lists 4K bitstreams (generated with VTM version 9.0, RA configuration of common test conditions) for which the bit rate is 40Mbps or less (maximum rate permitted by Level 5.1). Experiments were run on a computer featuring a single Xeon W processor with 8 cores (Skylake). “Real time” performance (60 frames per second or more) is achieved in all cases with the optimized decoder, while VTM processes between 2.8 and 4.6 frames per second.
Sequence | QP | Frames | Bit rate [Mbps] | VTM [s] | VTM [fps] | Optimized [fps] |
Tango | 22 | 294 | 27.9 | 93.792 | 3.13 | 95.4 |
FoodMarket | 22 | 300 | 17.0 | 89.978 | 3.33 | 103.8 |
Campfire | 27 | 300 | 30.7 | 93.977 | 3.19 | 110.8 |
CatRobot | 22 | 300 | 33.7 | 103.032 | 2.91 | 86.9 |
DaylightRoad | 27 | 300 | 10.3 | 81.181 | 3.70 | 109.3 |
ParkRunning | 32 | 300 | 22.2 | 107.255 | 2.80 | 92.1 |
DayStreet | 22 | 300 | 20.4 | 88.508 | 3.39 | 103.4 |
FlyingBirds | 22 | 300 | 15.9 | 69.801 | 4.30 | 121.6 |
PeopleInShoppingCenter | 22 | 300 | 11.6 | 65.853 | 4.56 | 127.2 |
SunsetBeach | 27 | 300 | 26.6 | 80.717 | 3.72 | 109.2 |
The three blocks with highest runtime consumption are inter prediction, ALF and deblocking. Runtime consumption of scaling and inverse transform is very small.
Though the optimized code runs significantly faster, the percentage of runtime consumed by individual tools (as measured by tool off tests) does not change significantly compared to the findings of AHG13.
Compared to a similarly optimized HEVC decoder, the runtime increase is around 1.5-2x.
It was asked if the runtime would be significantly slower for higher bit rates. This would probably be the case.
No specific action was necessary on this.
Profile/level specification (1)
See also section 4.5 for studies of coding tools and operation beyond the bit depth supported in VVC v1 profiles.
These contributions were first discussed in the Tuesday 30 June joint meeting.