Digital Signal Processing Reference
In-Depth Information
Cluster nodes A and B
in a hierarchical node
A
B
D
C
a
b
W
W'
{0,1}
{1,0}
D
C
D
C
CSDF version
SDF version
Fig. 9
Clustering of nodes A and B into an hierarchical node in (
a
)CSDFand(
b
)inSDF
a
b
1
1
1/99
1/99
current
frame
previous
frame
99
1
1
1
1
1
ME
EN
frME
EN
Fig. 10
A subgraph of an H.263 encoder graph (
a
)inSDFand(
b
)inFRDF
4
Other Decidable Dataflow Models
4.1
Fractional Rate Dataflow
The SDF model does not make any assumption on the data types of samples as
long as the same data types are used between two communicating nodes. To specify
multimedia applications or frame-based signal processing applications, it is natural
to use composite data types such as video frames or network packets. If a composite
data type is assumed, the buffer requirement for a single data sample can be huge,
which amounts to 176
144 pixels for a QCIF video frame for instance. Then
reducing the buffer requirement becomes more important than reducing the code
size when we make a schedule.
video frames. A QCIF video frame consists of 11
×
×
9 macroblocks whose size is
16
16 pixels. Node ME that performs motion estimation consumes the current and
the previous frames as inputs. Internally, the ME block divides the current frame
into 99 macroblocks and computes the motion vectors and the pixel differences from
the previous frame. And it produces 99 output samples at once where each output
×
