Graphics Reference
In-Depth Information
PU-Mode Pre-decision
CU-Layer High Complexity Mode Decision
64X64 CU
Intra Pred Dirs.
CU
0
HCMD
Cost
?
32X32 CU
HCMD
Cost
CU
1
CU
1
CU
1
CU
1
Inter PU Sizes & MVs
16X16 CU
?
HCMD
Cost
CU
2
CU
2
CU
2
CU
2
CU
2
CU
2
CU
2
CU
2
Fig. 11.16
Hardware-oriented two-step RDO algorithm flow
also used a mode pre-decision scheme at IME to reduce the computation for FME.
However, the BD rate increase for the fast RDO only method is quite high in HEVC.
If we cancel all the full RDO and use only fast RDO in HM, the BD rate increases
10
15 % in intra frames and may even increase to more than 40 % in inter frames.
Therefore, it seems that it is quite harmful to eliminate the full RDO completely
due to inaccurate prediction of rate-distortion cost. To keep the coding quality
while reducing cost for RDO process, hardware encoder should have a more limited
number of full RDO, but still keep the important decisions to full RDO.
Figure
11.16
shows the proposed two-step RDO algorithm flow. Since it is
expensive to use full RDO for all mode decisions, we use full RDO among the best
selected intra directions and inter prediction modes, and among CU depth levels.
For each CU depth level, the final direction and mode is decided only by fast RDO.
Thus, the number of modes is reduced to one per prediction type and CU depth level.
This mode pruning step occurs at intra prediction stage for intra modes and integer
motion estimation stage for inter modes. In intra prediction, the distortion cost is
SATD, and the rate cost is mode bits. In integer motion estimation, the distortion
cost is SAD, and the rate cost is motion vectors difference bits. In the next step,
more accurate costs for the selected modes are calculated by HCMD hardware,
which performs full RDO. The detail implementation of the HCMD hardware will
be discussed in Sect.
11.6.4
. After that, final mode is chosen accordingly. With the
two-step RDO algorithm, the number for full RDO that needs HCMD hardware is
decreased to 6, at the cost of 5:93 % BD-rate increase.
11.6.4
High Complexity Mode Decision
In the previous section, two-step RDO algorithm reduces the number of candidates
that require full RDO to 6. However, full RDO is still required to prevent large
quality drop. As a result, we still need efficient hardware design to take care of
