In the configuration files (in folder /hevc/cfg/), under heading #====Motion Search=====, in the parameter "FastSearch", it is given as 0 for Full search and 1 for EPZS, in the comments. But technically this comment is wrong, as the Fast search algorithm used in HM is TZSearch algorithm that is implemented in MVC. Basically these are the brief list of fast search algorithms in H.26x codecs. H.264/AVC JM - EPZS,UMHexS,SUMHexS H.264 MVC JMVM - TZSearch (or TZS algorithm) Now in HM, the same algorithm that is implemented in JMVM is implemented, which is TZS. But from most of the JCTVC meeting files (example JCTVC D600, JCTVC B300), this is understood as EPZS algorithm and which is not (from file /source/Lib/TLibEncoder/TEncSearch.cpp).

A newly introduced CABAC context (m_cCUSigCoeffGroupSCModel) is not being saved correctly in the encoder when WaveFrontSynchro is activated in the config. This causes decoder mismatches and crashes.

Although the minimum change is just to save the context correctly in TEncSbac::xCopyContextsFrom, the attached patches change both TEncSbac::xCopyContextsFrom and TDecSbac::xCopyContextsFrom, replacing the by-name context copies with a single memcpy. This will avoid future problems like this.

According to Text

( xT, yT ) = ( x0 + 2 << Log2CtbSizeY − 1, y0 − 1 ) (9-3)

when PicWidthInCtbsY is 2 and pic_width_in_luma_samples%CtbSizeY is not zero, (xT, yT) will be out of picture boundary, and availableFlagT should be FALSE.

But in HM-9.1, it only checks availability of top-right coding tree block, and availableFlagT will be TRUE.

This discrepancy should be fixed in HM or Text