I have used 0.9-hm branch r389, but it should apply to the most recent release as well. All changes are in TEncSearch::predInterSearch only.

1.Instead of:

#if GPB_SIMPLE_UNI
if ( pcCU->getSlice()->getSPS()->getUseLDC() )
{
...
}
else
{

xMotionEstimation ( pcCU, pcOrgYuv, iPartIdx, eRefPicList, &cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList [iRefIdxTemp], uiBitsTemp, uiCostTemp );

}
#else
...

, to bypass the L1 search we need this:

#if GPB_SIMPLE_UNI
if ( pcCU->getSlice()->getSPS()->getUseLDC() )
{
...
}
else
{

if ( iRefList && pcCU->getSlice()->getNoBackPredFlag() )
{

uiCostTemp = MAX_UINT; maybe not necessary
cMvTemp[1][iRefIdxTemp] = cMvTemp[0][iRefIdxTemp];

} else
xMotionEstimation ( pcCU, pcOrgYuv, iPartIdx, eRefPicList, &cMvPred[iRefList][iRefIdxTemp], iRefIdxTemp, cMvTemp[iRefList [iRefIdxTemp], uiBitsTemp, uiCostTemp );

}
#else
...

2. I also removed the check for LD when setting the first reference list for the bidirectional search. So at two places (within #if MS_NO_BACK_PRED_IN_B0) instead of:

if (pcCU->getSlice()->getSPS()->getUseLDC() && m_pcEncCfg->getUseFastEnc() && pcCU->getSlice()->getNoBackPredFlag() )

, I have:

if (m_pcEncCfg->getUseFastEnc() && pcCU->getSlice()->getNoBackPredFlag() )

My understanding is that the results are slightly different now because the L1 MV is searched for in the bidirectional part of the ME, and not during the unidirectional one as before. Without fix we have:

L0 search -> L0 MV
L1 search -> L1 MV
Init BI MV with L0 MV and L1 MV
Refine L0 MV for BI

An now we have:

L0 search -> L0 MV
Init L1 MV with L0 MV
Init BI MV with L0 MV and L1 MV
Refine L1 MV for BI