Graphics Reference
In-Depth Information
Full RDO Pass
Intra
Prediction
Fast RDO
(30+ modes)
IQ
IT
SSD
Final
Mode
Decision
T
Q
CABAC
Rate
Motion
Estimation
T/Q: Transform/Quantization
IT/IQ: Inverse Transform / Quantization
Fig. 11.14
RDO algorithm flow in HM
Firstly, fast RDO is done. To save computation overhead, fast RDO selects
several candidates among intra prediction directions and inter prediction motion
vectors and modes for each depth level. In fast RDO, the rate is determined by
mode bits and motion vectors, and the distortion is calculated by the sum of absolute
difference (SAD) or the sum of absolute transformed difference (SATD). Although
fast RDO is not accurate, it is still able to prune the less probable cases. Fast RDO
selects totally more than 30 modes depending on encoder configurations. Then, the
full RDO costs are estimated and compared for these modes. In full RDO process, all
the residues are transformed, quantized, inverse quantized, and inverse transformed
to produce the reconstructed differences. Distortion is then calculated by the sum of
squared difference (SSD). The prediction information and residue coefficients will
go through CABAC bit estimator to obtain bit rate if estimated mode is selected.
After that, final decision between the modes is made by Lagrangian cost with SSD
distortion and estimated CABAC bit rate to optimize the trade-off.
11.6.2
Proposed Hardware RDO Mode Decision Pipeline
In hardware, we also use a hardware-oriented two-step RDO algorithm for mode
decision. Figure
11.15
shows the overall RDO mode decision hardware architecture.
RDO mode decision requires several major functional units to cooperate. Thus,
several CU-level pipeline stages are shown. In the first step, mode pruning is
done in intra and IME stages. FME refines all the modes selected by inter motion
estimation. After that, full RDO is performed for each mode. A High Complexity
Mode Decision (HCMD) hardware consisting of a bit rate estimator and a SSD
cost unit is used for each mode that needs full RDO. The final mode is decided
by comparing the resulting costs from HCMD hardware in all selected candidates.
After that, the context state update for bit rate estimator is performed according to
the final modes. More details on the HCMD hardware are provided in Sect.
11.6.4
.
