Opened 11 years ago Closed 11 years ago #1262 closed defect (fixed)Reference decoder discards chroma CBFs when decoding 4:2:2 with log2_min_transform_block_size_minus2>0
Description
In TDecSbac::parseQtCbf(), there is this piece of code: if ((width != height) && (lowestLevel || !canQuadSplit)) //if sub-TUs are present { const UInt subTUDepth = lowestTUDepth + 1; const UInt partIdxesPerSubTU = rTu.GetAbsPartIdxNumParts(compID) >> 1;
When the total number of partitions in a CU are calculated, the sub-TU partitions for 4:2:2 chroma are not taken into account. This means that when decoding, for example, an 8x8 TU when log2_min_transform_block_size_minus2==1, rTu.GetAbsPartIdxNumParts(compID) will come out as 1, which will give partIdxesPerSubTU=0, causing the decoded CBF information for the sub-TUs to be discarded (stored across a block of length zero).
This doesn't happen when log2_min_transform_block_size_minus2==0 as the 4x4 chroma sub-TUs are then at least one layer above the lowest luma layer, so sufficient partition indices are available to store the data. Change History (4)comment:1 Changed 11 years ago by DefaultCC Plugin
comment:2 Changed 11 years ago by karlsharmancomment:3 Changed 11 years ago by jackh
I can confirm that this is fixed. comment:4 Changed 11 years ago by karlsharman
Note: See
TracTickets for help on using
tickets. | This list contains all users that will be notified about changes made to this ticket. These roles will be notified: Reporter, Owner, Subscriber, Participant
|
A fix was added in r3869. This adds an additional depth to the structures (when 4:2:2 and minimum luma TU size is not 4x4), thereby giving sufficient parameter storage for the 4:2:2 square-sub-TU split.
g_maxCUDepth and g_uiAddCUDepth are 1 greater for these cases, as controlled by the function getMaxCUDepthOffset in TComChromaFormat.h
Example that I found to cause the above situation:
./bin/TAppEncoderStatic -c cfg/encoder_intra_main_rext.cfg
Please would the reporter check this solves their particular configurations.
Also linked to #1259.