Back to Search Document details
16th Meeting: Geneva, October 2019 2019-10-04 14:07
Non-CE6: Removal of LFNST for 4x4, 4xN and Nx4 blocks
Authors: H. E. Egilmez Qualcomm T. Hsieh Qualcomm Vadim Seregin Qualcomm Marta Karczewicz Qualcomm
Abstract
This document proposes to disable LFNST for 4x4, 4xN and Nx4 blocks and reduce latency, which is critical for small blocks, and it reduces 2KB of memory. The experimental results show that the proposed simplification leads to 0.16% AI and 0.10% RA luma BD-rate losses over VTM-6.0, and the encoder runtime is reduced about 10% for AI coding.
Meeting Notes
JVET-P0634 Non-CE6: Removal of LFNST for 4x4, 4xN and Nx4 blocks [H. E. Egilmez, T. Hsieh, V. Seregin, M. Karczewicz (Qualcomm)] [late]

This document proposes to disable LFNST for 4x4, 4xN and Nx4 blocks and reduce latency, which is critical for small blocks, and it reduces 2KB of memory. The experimental results show that the proposed simplification leads to 0.16% AI and 0.10% RA luma BD-rate losses over VTM-6.0, and the encoder runtime is reduced about 10% for AI coding.

Latency is not critical at decoder, as the residual can be reconstructed separately from the prediction. An encoder that has latency problems might better disable LFNST. The loss is too large to justify the change.

References:
JVET-L1026 JVET-M0026 JVET-N0026 JVET-N0193 JVET-N1010 JVET-O0026
Cited:
JVET-P0026 JVET-P0987 X
PATENTS:
KR102661314B1 0.34 2024-04-30 KR102699362B1 0.32 2024-08-28 KR102396112B1 0.30 2022-05-10 AU2023202526A1 0.28 2023-05-11 KR20220081999A 0.26 2022-06-16 KR102806831B1 0.24 2025-05-14 JP7754907B2 0.22 2025-10-15 CN113812154B 0.20 2023-01-10
ARCHIVE v2 :
JVET-P0634.pptx JVET-P0634_v1.docx Test.XLSM
Decisions
Latency is not critical at decoder, as the residual can be reconstructed separately from the prediction. An encoder that has latency problems might better disable LFNST. The loss is too large to justify the change.
Citation