Bug in Table 131

[peterchuang]

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs >= 0 && !slice_ts_residual_coding_disabled_flag

should be

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs > 0 && !slice_ts_residual_coding_disabled_flag

the ">=" should be ">"

[sunmi.yoo]

Currently, checking ccbs is conducted at a pass level, which means it does not decrease syntax element by syntax element. Therefore, the bug in Table 131 can be fixed from :

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 0 | | RemCcbs == 0 | | slice_ts_residual_coding_disabled_flag

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs >= 0 && !slice_ts_residual_coding_disabled_flag

to

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 0 | | RemCcbs < 4 | | slice_ts_residual_coding_disabled_flag

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs >= 4 && !slice_ts_residual_coding_disabled_flag

[wangyang.cs]

In residual_ts_coding(), coeff_sign_flag is context coded in the first scan pass. RemCcbs could be equal to 3 in the first scan pass since sig_coeff_flag may be decoded and RemCcbs will minus 1.
coeff_sign_flag is bypass coded in the remainder scan pass in residual_ts_coding(). RemCcbs could be less than or equal to 3.
Only changing Table 131 is not enough to distinguish whether coeff_sign_flag is bypass or context coded in residual_ts_coding(). To fix this problem, RemCcbs could be set equal to 0 before the remainder scan pass in residual_ts_coding().

My suggestion is as follows:
In JVET-R2001-vA
7.3.10.11 Residual coding syntax

residual_ts_coding( x0, y0, log2TbWidth, log2TbHeight, cIdx ) {	Descriptor
...
/* remainder scan pass */
RemCcbs = 0
for( n = 0; n <= numSbCoeff − 1; n++ ) {
xC = ( xS << log2SbW ) + DiagScanOrder[ log2SbW ][ log2SbH ][ n ][ 0 ]
yC = ( yS << log2SbH ) + DiagScanOrder[ log2SbW ][ log2SbH ][ n ][ 1 ]

Table 131 could be changed from

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 0 | | RemCcbs == 0 | | sh_ts_residual_coding_disabled_flag
coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs >= 0 && ! sh_ts_residual_coding_disabled_flag

to

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 0 | | RemCcbs == 0 | | sh_ts_residual_coding_disabled_flag
coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs > 0 && ! sh_ts_residual_coding_disabled_flag

[sunmi.yoo]

As wangyang.cs commented, my previous suggestion cannot fix the bug in the case when RemCcbs is 4 before coeff_sig_flag is parsed and RemCcb is decremented after coeff_sig_flag.

However, if the spec is modified as the wangyang's suggestion, only coeff_sign_flag in the first CG of a TU can be context coded, otherwise always coeff_sign_flag is bypassed since RemCcbs is set to 0 in the first CG. Another suggestion in Table 131 (+ no modification in 7.3.10.11) is as follows :

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 0 | | ( RemCcbs < 3 && lastScanPosPass1 = -1 ) | | sh_ts_residual_coding_disabled_flag

coeff_sign_flag[ ]
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs >= 3 && ! sh_ts_residual_coding_disabled_flag

[m.sarwer]

I also agree that the conditions in Table 131 is not correct in some cases. However, I am not sure how the solution proposed here will work in following case:

Assume, there are 4 CG and after coding 3 CG, the RemCcbs is equal to 3. Based on the above solution, following condition is true which means coeff_sign_flag is context coded.

transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && RemCcbs >= 3 && ! sh_ts_residual_coding_disabled_flag

But actually in VTM implementation and the intention in the above mentioned example, coeff_sign_flag is by-pass coded.

Below is an alternative solution to fix that. Following are the required changes.

Changes in Table 131:

Bypass condition:
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 0 | | n > lastScanPosPass1 | | sh_ts_residual_coding_disabled_flag

Condition for context coded:
transform_skip_flag[ x0 ][ y0 ][ cIdx ] = = 1 && n <= lastScanPosPass1 && !sh_ts_residual_coding_disabled_flag

Changes in the syntax table:

The variable “ lastScanPosPass1 = n” need to define at the beginning of the first pass loop instead of end of the loop.

[sunmi.yoo]

@m.sarwer,

I also agree with your point that my second suggestion still does not solve the issue.

However, in your solution, it seems that there is no chance for coeff_sign_flag[] to be context coded in every first non-zero coefficient in the CG, because lastScanPosPass1 is initialized to be -1 in every CG and updated after the first non-zero coefficient is parsed.

We'd better look for possible solutions.

[sunmi.yoo]

@m.sarwer,

Your suggestion seems to solve the issue.

I missed your last sentence for the changes in the syntax table

The variable “ lastScanPosPass1 = n” need to define at the beginning of the first pass loop instead of end of the loop.

[bbross]

This will be fixed in JVET-S2001-v7 with the adoption of JVET-S0215.