JVET-AD0187 Non-EE2: On large NSPT [M. Koo, J. Zhao, J. Lim, S. Kim (LGE)]
In the 29th JVET meeting, Non-Separable Primary Transform (NSPT) was incorporated into ECM-8.0, where the NSPT is applied to 4x4, 4x8/8x4, 4x16/16x4, 8x8, and 8x16/16x8 transform blocks instead of the combination of DCT-2 and LFNST. This contribution introduces additional large NSPT for Nx32/32xN (N = 4 and 8) blocks of which matrix dimensions are 36x128 (N = 4) and 48x256 (N = 8), respectively. The number of transform sets and kernels per set are not changed for the proposed large NSPT. The BD-rate changes and encoding/decoding times for AI configuration are -0.09%/-0.11%/-0.11% for Y/U/V and 102%/100%, respectively.
Why is encoding time increased?
Gain is relatively small, compared to previous NSPT adoptions, and additional larger kernels are introduced.
Kernels were trained the same way as the existing ones
Some interest was expressed in the proposal.
It was agreed to investigate this in an EE, reduction of coding time should be studied, worst case complexity, and memory needs for the additional kernels.