Graphics Reference
In-Depth Information
0
12 3
4
3
0
1
2
01 2
0
1
2
3
0
1
2
0
1
0
Fig. 10.2
Split system pipeline to address variable DRAM latency. Within each group, variable-
sized pipeline block-level pipelining is used
The entropy decoder must send residue coefficients and transform information
such as quantization parameter and TU size to the inverse transform block. As
residue coefficients use 16-bit precision, 12 kB of SRAM is needed for luma and
chroma coefficients of one VPB. For full pipelining, storage for two VPBs is
needed so that entropy decoder can write coefficients and inverse transform can
read coefficients of the previous VPB simultaneously. Thus, VPB pipelining would
need 24 kB of SRAM. But this can be avoided by using the fact that the largest TU
size is 32
32 (a 64
64 CU must split its transform quadtree at least once). Hence,
it is possible to use a 2-TU buffer instead. The entropy decoder writes to one TU
while inverse transform reads from the previous TU. This buffer requires only 4 kB,
thus saving 20 kB of SRAM.
In the first pipeline group, a line buffer is used by entropy decoder for storing
prediction information of upper row VPBs. In the second pipeline group, the 9-
bit residues are passed from inverse transform to prediction using two VPB-sized
SRAMs in ping-pong configuration. (Inverse transform writes one VPB to one
SRAM while prediction reads the previous VPB from the other SRAM. When
both modules are finished processing their respective VPBs, the two SRAMs switch
roles.) Prediction, in-loop filters and reconstruction DMA communicate using three
VPB-sized SRAMs in a rotating buffer configuration as shown in Fig.
10.3
. Another
line buffer is used to communicate pixels and parameters across VPB rows. The line
buffer must store:
four luma and two chroma rows (pre-deblocking) for deblocking filter. Of these,
one luma and one chroma rows are also used as top neighbor pixels for intra
prediction.
one luma and one chroma rows (post-deblocking) for SAO filter
