Back to Search Document details
29th Meeting: by teleconference, DE, January 2023 2023-01-15 10:00
On worst case decoder complexity of JVET-AC0130
Authors: Jonathan Gan OPPO Yue Yu OPPO
Abstract
This contribution proposes reducing the number of basis vectors used in the NSPT of JVET-AC0130 for 8x16 and 16x8 block sizes from 40 to 32, which is asserted to maintain the worst-case decoder complexity at the inverse transform stage to 32 multiplications per sample. The reported results are AI -0.43% -0.57% -0.53% enc 103% dec 100%.
Meeting Notes
JVET-AC0332 On worst case decoder complexity of JVET-AC0130 [J. Gan, Y. Yu (OPPO)] [late]

This contribution proposes reducing the number of basis vectors used in the NSPT of JVET-AC0130 for 8x16 and 16x8 block sizes from 40 to 32, which is asserted to maintain the worst-case decoder complexity at the inverse transform stage to 32 multiplications per sample. The reported results are AI -0.43% -0.57% -0.53% enc 103% dec 100%.

The results confirm that the 8x16 and 16x8 transforms are beneficial in terms of gain. It is further shown that the reduction of the number of basis vectors practically has no loss.

It was pointed out that the DCT 256x256, and MST in general, may be more costly, implementation-wise. So, the contribution would not reduce the worst case.

No need to take action on this was identified.

A crosscheck was reported in JVET-AC0347.

References:
JVET-AC0130
Cited:
JVET-AC0347 X
ARCHIVE :
JVET-AC0332.docx JVET-AC0332.xlsm NSPT-add16x4_4x16.xlsm NSPT-add16x8_8x16.xlsm NSPT-test4.2a-SIMD.xlsm NSPT-test4.2b-SIMD.xlsm
Decisions
A crosscheck was reported in JVET-AC0347.
Citation