MV rounding for IBC AMVP mode

[chhuanb]

In clause 8.6.2.2, rounding process for IBC MVP was not performed.

[bbross]

Anyone willing to suggest a spec fix that is aligned with what VTM is doing?
I assume the MVs are rounded to integer sample accuracy.
However, it needs to be clarified, where the rounding happens, e.g.
a) each MVP candidate when inserted in the list similar to inter AMVP
b) only the final MVP, i.e. mvL

[axink]

Hi Ben, all,

Currently, the ibc merge candidate and ibc MVP candidate share the same predictor list. So during list construction, no rounding operation should be added (merge mode does not need MV rounding). The MV rounding operates for the selected MV candidate (final MVP) in the MVP mode. This is also what the VTM is doing.

Best,
Xiaozhong

A suggested fix is provided as follows:

In 8.6.2.1,

...
– When general_merge_flag[ xCb ][ yCb ] is equal to 0, the following applies:

---

added text starts

1. The rounding process for motion vectors as specified in clause 8.5.2.14 is invoked with mvX set equal to mvL, rightShift set equal to MvShift + 2, and leftShift set equal to MvShift + 2 as inputs and the rounded mvL as output.

added text ends

---

2. The variable mvd is derived as follows:

...

[bbross]

Fixed in D5 v7