Digital Signal Processing Reference
In-Depth Information
a
b
v
1
v
2
Fig. 9
Splitting of the input array into two parts. (
a
) Token transport.
Gray shaded rectangles
belong to the extended border. (
b
)WSDFgraph
Figure
9
depicts the corresponding scenario. It exactly matches the scenario
from actor
A
's point of view, the situation has not changed. It still produces the same
tokens in both scenarios. However, actor
B
groups these tokens into two subarrays,
instead of one. Each of these subarrays is extended by its own extended border.
In particular, the sampling with sliding windows restarts for each sub-array at the
upper left corner. Note furthermore that the amount of tokens produced by actor
B
is the same for both scenarios, while their values will differ because of the different
border processing. Note in addition that while the tokens generated by actor
B
will
originate from two different input subarrays, actor
C
unifies the tokens again into
one array. Such a behavior is again useful in applications like JPEG 2000, where the
different tiles are combined into one output file.
In order to support such scenarios, WSDF edges are annotated by a vector
v
.It
defines how many data elements are grouped to one subarray, the so called
virtual
token
. Note that it does not impact the way how the source of an edge is behaving.
Only the edge sink is concerned by performing the corresponding border processing
and restarting the sampling at each upper left corner of each input subarray. This
concept by the way leads to a balance equation with similar properties than for
4.2.4
Graphs with Feedback Loops and Non-Rectangular Windows
Similar to cyclo-static dataflow graphs, feedback loops can be handled in WSDF
either by placing so called initial tokens on the edges, or by carefully designing the
