JVET-AF0201 AHG6: On ECM SW memory consumption [R. Chernyak, M. Xu, S. Liu (Tencent), Y. Yasugi, T. Ikai (Sharp)]
This document provides a memory consumption comparison between ECM9, ECM10 and one of the recent versions of ECM master branch. It is observed that significant improvement of the ECM memory consumption was performed after ECM9 release and worst case of encoder time memory consumption was reduced for more than 50%. This document also proposes to extend the current ECM encoder memory measurement mechanism to the decoder side.
This contribution pointed out an increase in SAO-related encoder memory consumption between ECM-9.1 and ECM-10.0.
m_cCuEncoder | m_cEncAlf | m_cEnc_Sao | Other | Total | |
ECM9 | 0.7 | 21 | 0.5 | 0.2 | 22.4 |
ECM10 | 0.7 | 6.9 | 2.3 | 0.2 | 10.1 |
ECM10-9e651 | 0.2 | 6.9 | 2.3 | 0.1 | 9.5 |
It was asked what might have caused the big difference in SAO-related memory consumption. The proponent has not looked into this issue.
Another expert commented that this could be because more edge classifiers were adopted into CCSAO that caused the memory increase. The proponent of more edge classifiers (JVET-AE0151) reported a much smaller memory consumption increase based on internal test, and the increase observed by the JVET-AE0151 proponent was more in line with the numbers reported independently in JVET-AF0090.
It was commented that encoder memory consumption is a much bigger issue than decoder memory consumption. However, several experts commented that it would be a good practice to also track decoder memory consumption.
It was agreed to update the CTC document JVET-AF2017 to include an xls template that contains decoder memory comsuption information.