Graphics Reference
In-Depth Information
b
a
Fig. 10.19
Tight feedback loop in intra prediction due to dependency between neighbors.
(
a
) Intra-prediction dependency between neighboring pixel blocks. (
b
) Dependency results in a
tight feedback loop
The key operations in intra-prediction are:
1. Read neighboring pixels and perform padding for unavailable pixels
2. Reference preparation: filter neighboring pixels to obtain intra reference pixels
and extend the top-left reference pixels for angular modes
3. Prediction: bilinear interpolation for angular and planar modes, and pixel copy
for DC, horizontal and vertical modes
When the current block of pixels is predicted, its residues need to be immediately
added so that it can be used as neighboring pixels for the next block. This results
in a tight feedback loop for intra-prediction as shown in Fig.
10.19
.Asaresultof
this feedback loop, it is not possible to pipeline the above three operations, which
increases the throughput requirement from these blocks. It should be noted that the
feedback loop operates at a TU granularity and not a PU granularity. For example,
for a 16
16 CU with a 2N
2N intra partition (i.e. a single 16
16 PU) and a
residue quad tree (RQT) of four 8
8 TUs, the 8
8 blocks must be predicted serially
and the intra neighboring pixels must be updated after every block's prediction and
reconstruction.
This dependency also has implications for the top-level pipelining—in order to
keep inverse transform and prediction decoupled, the inverse transform must be
performed one pipeline granularity before prediction.
The 35 intra prediction modes in HEVC are well designed to reduce com-
plexity. The planar mode is much simpler than the one in H.264/AVC, and the
33 angular modes are also well organized to avoid increasing the complexity
when increasing the angular precision. However, the larger TU sizes increase the
hardware complexity due to larger pipeline and reference buffers. In H.264/AVC,
