Embedded Software for SoC - Grupo de Mecatrônica EESC/USP

134 Chapter 10
Table 10-1. Execution time of H.263 encoder system.
Architecture
Point-to-point interconnection
Execution
cycle
Runtime
(sec)
Accuracy
(%)
Shared bus
Execution
cycle
Runtime
(sec)
Accuracy
(%)
Cosimulation
ILP
w/o dyn. SW
w/ dyn. SW
Heuristic
5,031,708
4,684,812
4,979,535
4,979,535
4348
1.71
2.23
4.11
100
93.10
98.96
98.96
5,151,564
4,804,668
5,099,391
5,099,391
29412
10.17
14.37
4.23
100
93.27
98.98
98.98
could enhance the accuracy from 93.10% to 98.96% for the point-to-point
interconnection architecture. Similar enhancement was obtained for the other
architecture. We could not obtain 100% accuracy because we did not consider
some dynamic behaviors of task VLC and OS scheduler. In this example, the
schedule obtained by the heuristic algorithm was the same as that obtained
by ILP. That is why we obtained the same accuracy. Table 11-1 shows also
the runtimes of cosimulation, ILP, and heuristic algorithm in seconds. In case
of point-to-point interconnection, heuristic runtime is larger than that of ILP.
That is because the number of nodes in ETG is so small that the complexity
depends on constants rather than the number of nodes. Note also that the
cosimulation was performed for the schedule obtained by ILP and the runtime
does not include the scheduling time.
Then we experimented with larger examples including JPEG and IS-95
CDMA. Table 10-2 shows a comparison of system execution times for different
sizes of physical buffers. The third column shows the numbers of nodes
in ETG and the last column shows the errors obtained by comparing with
ILP optimal scheduling.
As Table 10-2 shows, as the size of physical buffer gets smaller, the system
execution time increases. It is due to the synchronization delay caused by the
conflicts in accessing the shared buffer. As the table shows, the ILP solutions
without dynamic SW behavior give the same number of system
execution cycles for three different sizes of shared buffer. This implies that
without considering the dynamic SW behavior, such conflicts are not
accounted for during the scheduling. As shown in the table, the ILP without
the dynamic SW behavior yields up to 15.13% error compared to the ILP with
the dynamic SW behavior. Thus, to achieve accurate scheduling of communication
and task execution, the dynamic SW behavior needs to be incorporated
into the scheduling.