Graphics Reference
In-Depth Information
signal. To save on the transform decision cost, one of the options in HM is to use the
largest available transform size. Some previous works [
15
] propose fast transform
decision by early split termination. Owing to the fact that 32
32 TU is less selected
than 8
8 TU and 16
16 TU, Teng [
37
] proposes a split-and-merge process. While
this method works for software, it is not suitable for hardware implementation with
variable hardware resource requirements and non-predictable computation cycles.
For hardware architecture, a reusable architecture for various TU size is proposed in
[
31
]. A unified engine for forward and inverse transform architecture for HEVC is
shown in [
2
,
45
]. The distribution of transform residuals has strong relationship with
the homogeneity of the predicted results. As a result, we choose to use the largest
possible transform size inside PU boundary. With this fixed decision, the hardware
cost for transform is reduced at the cost of BD-rate increase of 3:02 %inlowdelay
P configuration.
11.6
Rate Distortion-Optimization
Rate distortion optimization (RDO) is an important method in video coding.
It is based on the Lagrange optimization technique. With the proper choice of
parameters, optimal trade-off between rate and distortion can be achieved. In
HEVC, the combinations of prediction modes are much more complex than that
in H.264/AVC. Low precision fast RDO is used in previous H.264/AVC encoders
based on rough estimation of bit rate and distortion. The decision quality, however,
is quite low. A more precise mode decision is thus required to maximize the coding
gain. In HM, a full RDO method based on CABAC bit rate estimation and SSD
distortion cost is used. However, the cost is very high in real-time video encoder
hardware. In this section, we discuss the design of cost efficient RDO hardware
with high decision quality.
11.6.1
Rate-Distortion Optimized Mode Decision in HM
The mode decision process of HM is shown in Fig.
11.14
.HMusesatwo-stepRDO
algorithm for mode decision. Firstly, fast RDO is used for early termination, and
then full RDO is used for final decision. To have better mode decision quality and
improve coding gain, full RDO is applied among the following aspects:
The directions and modes chosen by fast RDO.
The best selected intra directions and inter prediction modes.
CU depth levels.
