JVET-AQ0150 [AHG16] Improvements on Hardware Implementation Complexity (HIC) Software [K. Naser, F. Le Léannec, F. Galpin, F. de Lagrange (InterDigital), Y. Zhao, E. Alshina (Huawei)]
This contribution proposes multiple improvements to the HIC software, following the recommendations by hardware experts comments from last meeting. Specifically, it proposes the following:
- The RDO budget of intra coding modes in I frame is increased by reusing the resources for inter mode RDOs, where intra modes can additionally use half of the inter RDOs, according to the suggestion provided in JVET-AQ0045.
- Defining a per-CU RDO budget related to CU sizes, such that smaller CU’s are more restricted in terms of the number of RDOs (vice-versa).
It is reported that the new performance of constraint encoder (reduced runtime environment, GOP8, 1 reference frame and 10 RDOs for large CUs & 5 RDOs for small ones) compared to VTM-24.0 anchor is:
20.86% 31.03% 31.30% 13% 104%
HIC-1.0 SW (10 RDOs for all CUs) vs. VTM-24.0 anchor is:
20.11% 29.79% 30.29% 14% 105%
And HIC-1.0 SW (5 RDOs for all CUs) vs. VTM-24.0 anchor is:
23.27% 36.43% 36.17% 12% 105%
Test1: Impact of using half max inter RDOs in intra frames
Anchor: HIC-1.0 with GOP8 + 1 reference frame + RR3 +10 IntraRDOs + 10 InterRDOs
Test1: Anchor + Intra modes in I Frames use additionally half inter RDOs (15 RDOs)
|
| Random access Main10 |
|
| |
|
| Over VTM-23.9 |
|
| |
Y | U | V | EncT | DecT | |
Class A1 | -0.19% | -0.09% | -0.23% | 100% | 99% |
Class A2 | -0.28% | 0.02% | -0.15% | 100% | 100% |
Class B | -0.30% | 0.05% | -0.53% | 101% | 100% |
Class C | -0.41% | -0.11% | -0.46% | 102% | 100% |
Class E |
|
|
|
| |
Overall | -0.30% | -0.03% | -0.38% | 101% | 100% |
Class D | -0.25% | 1.25% | 0.03% | 102% | 102% |
Class F | -1.78% | -2.27% | -1.98% | 102% | 101% |
It can be concluded that 0.3% coding gain can be obtained without sacrificing the encoder time.
Test2: Impact of using half max inter RDOs in intra frames and further restricting RDOs for smaller CUs
Anchor: same as in Test1
Test2: Test1 + 5 to 10 Max RDOs as in Table 1
|
| Random access Main10 |
|
| |
|
| Over VTM-23.9 |
|
| |
Y | U | V | EncT | DecT | |
Class A1 | 0.71% | 0.81% | 0.38% | 96% | 100% |
Class A2 | 0.31% | 0.79% | 0.49% | 96% | 100% |
Class B | 0.53% | 0.65% | 0.60% | 96% | 100% |
Class C | 0.91% | 1.22% | 1.37% | 94% | 99% |
Class E |
|
|
|
| |
Overall | 0.62% | 0.86% | 0.74% | 95% | 99% |
Class D | 1.45% | 3.32% | 2.30% | 94% | 102% |
Class F | 1.74% | 1.39% | 0.95% | 95% | 99% |
Clearly, this comes with a coding loss, but also with reduction of encoding time. It can also be observed that there is some larger loss for small resolution sequences, as fewer RDOs are used.
Test3: VTM as anchor
Anchor: VTM-24.0
Test3a: HIC-1.0 with GOP8 + 1 reference frame + RR3 +10 IntraRDOs + 10 InterRDOs
Test3b: HIC-1.0 with GOP8 + 1 reference frame + RR3 +5 IntraRDOs + 5 InterRDOs
Test3c: HIC-1.0 with GOP8 + 1 reference frame + RR3 +5 to 10 IntraRDOs + 5 to 10 InterRDOs + intra frame further use half max inter RDO
Test3a: HIC-1.0 with GOP8 + 1 reference frame + RR3 +10 IntraRDOs + 10 InterRDOs
|
| Random access Main10 |
|
| |
|
| Over VTM-24.0 |
|
| |
Y | U | V | EncT | DecT | |
Class A1 | 15.26% | 24.48% | 26.66% | 13% | 106% |
Class A2 | 22.32% | 34.66% | 32.08% | 14% | 104% |
Class B | 22.13% | 32.52% | 34.50% | 14% | 105% |
Class C | 19.55% | 26.71% | 26.41% | 13% | 103% |
Class E |
|
|
|
| |
Overall | 20.11% | 29.79% | 30.29% | 14% | 105% |
Class D | 15.13% | 19.01% | 23.39% | 13% | 93% |
Class F | 24.18% | 26.03% | 28.21% | 20% | 104% |
Test3b: HIC-1.0 with GOP8 + 1 reference frame + RR3 +5 IntraRDOs + 5 InterRDOs
|
| Random access Main10 |
|
| |
|
| Over VTM-24.0 |
|
| |
Y | U | V | EncT | DecT | |
Class A1 | 18.71% | 29.94% | 31.52% | 11% | 107% |
Class A2 | 25.08% | 41.63% | 37.20% | 13% | 104% |
Class B | 25.04% | 39.64% | 40.99% | 13% | 106% |
Class C | 23.13% | 33.38% | 32.85% | 11% | 103% |
Class E |
|
|
|
| |
Overall | 23.27% | 36.43% | 36.17% | 12% | 105% |
Class D | 18.61% | 27.52% | 30.84% | 11% | 94% |
Class F | 32.60% | 39.56% | 41.78% | 17% | 103% |
Test3c: HIC-1.0 with GOP8 + 1 reference frame + RR3 +5 to 10 IntraRDOs + 5 to 10 InterRDOs + intra frame further use half max inter RDO
|
| Random access Main10 |
|
| |
|
| Over VTM-24.0 |
|
| |
Y | U | V | EncT | DecT | |
Class A1 | 16.08% | 25.55% | 27.18% | 12% | 105% |
Class A2 | 22.67% | 35.78% | 32.73% | 14% | 104% |
Class B | 22.80% | 33.60% | 35.27% | 14% | 105% |
Class C | 20.65% | 28.37% | 28.35% | 12% | 102% |
Class E |
|
|
|
| |
Overall | 20.86% | 31.03% | 31.30% | 13% | 104% |
Class D | 16.85% | 23.00% | 26.54% | 12% | 94% |
Class F | 26.29% | 28.09% | 29.65% | 19% | 102% |
It is noted that modified SW (Test3c) achieves almost the same coding performance as HC SW with 10 RDOs, while not reducing the maximum RDOs for the small CUs. It also save ~2.5% of BD rate compared to test2, where all CU’s are restricted to 5 RDOs. Therefore, it is recommended to use the Test3c configuration for the HIC software.
It was suggested to investigate whether the macros from the software could be used to measure the RDO count in VTM in the context of the complexity reporting template for the CfP (both for anchor and faster encoding methods). This could also be provided to proponents as an example how they could measure the RDO count.
Decision(SW): include JVET-AQ0150 software in the AHG16 software