At the end of function TEncSearch::xMotionEstimation, we have the following line

ruiCost = (UInt)( floor( fWeight * ( (Double)ruiCost - (Double)m_pcRdCost->getCost( uiMvBits ) ) ) + (Double)m_pcRdCost->getCost( ruiBits ) );

For P and B frames, fWeight=1 and 0.5, respectively. When we calculate the distortion (SAD/HAD) for B frames in motion estimation, we already used the average prediction of two hypotheses, such as in TComRdCost::xGetSADs32(). Then fWeight should also be set to 1 for B frames. To keep the original design (fWeight=0.5 for B frames) unchanged, during the calculation of sub-pixel distortion such as TComRdCost::xCalcHADs8x8, a one-bit left shift is applied after the difference calculation, i.e.,

pred = xClip( (piCur[i*iStep] + piRef[i] + bRound) >> 1 );
diff[k+i] = (piOrg[i] - pred) << 1;

which compensates the setting of fWeight=0.5. Such an implementation (fWeight=0.5 + one-bit left shift) is correct. However, the code is confused and difficult to understand. We hope it can be improved to be more clear. For example, set fWeight=1 for B frames while removing the left shift in functions like TComRdCost::xCalcHADs8x8.