JVET-L0137 CE1-related: Minimum block size restriction [J. Choi, J. Heo, S. Yoo, L. Li, J. Choi, J. Lim, S. Kim (LGE)]
In VTM2.0.1, a CU can be split into 4x4 luma block in both intra and inter slices. A size of minimum block (CU or PU) plays an important role for worst case complexity in aspects of memory bandwidth and throughput in hardware architecture. In addition, the coding performance impact of small block coding is getting less significant but requires higher worst case complexity as input video resolution increases. Therefore, this contribution suggests minimum block size restriction methods as follows:
- Method 1: chroma 2x2 block is restricted in I slice, and luma 4x4 and chroma 2x2 are restricted in P and B slice.
- Method 2: luma 4x4 block and chroma 2x2 block are restricted in all slices.
The experimental results for Method 1 reportedly show luma BD-rate losses of 0.00%, 0.22% and 0.18% compared to VTM2.0.1 in AI, RA and LB configurations, respectively. The experimental results for Method 1 reportedly show luma BD-rate losses of 0.46%, 0.36% and 0.29% compared to VTM2.0.1 in AI, RA and LB configurations, respectively. In addition, the BD-rate changes for Method 2 were found to be reduced to 0.19%, 0.17% and 0.12% for UHD and HD test sequences.
An interesting aspect is that the losses are less severe in case of high resolution, where the memory bandwidth is a more severe issue. This would not resolve a case where the same decoder would either decode UHD or four HD streams simultaneously (as the constraint would only apply for UHD). Applying at HD and above might be useful.
It is mentioned that the contribution likely did not consider restricting affine and other subblock related motion comp (as these are implicitly using 4x4) – disabling them would result in higher loss.
A potential solution for limiting memory bandwidth problems with subblock MC tools and 4x4 block size would be an encoder restriction that would not allow large variation. (contribution JVET-L0396 is related to this).