Embedded Software for SoC - Grupo de Mecatrônica EESC/USP

128 Chapter 10
3. ON-CHIP COMMUNICATION SCHEDULING AND TIMING
ANALYSIS
For an accurate timing estimation of on-chip communication for design space
exploration, we actually perform optimal scheduling of the communication.
For the accuracy, we consider the dynamic behavior of SW. We also consider
the effect of buffer sharing on the communication scheduling.
3.1. Problem definition and notations
Problem
Given a task graph, a target architecture (with its parameters, e.g. maximum
physical buffer sizes on shared memory components), and a mapping of tasks
and communications on the target architecture (including physical buffer
sharing), the problem is to find a scheduling of task execution and communication
which yields the minimum execution time of the task graph on the
target architecture.
To solve the problem, we extend the task graph to an extended task graph
that contains detailed communication operations (in our terms, communication
nodes). Then we apply an ILP formulation or heuristic algorithm to
schedule tasks and communications. For the on-chip communication, we
assume on-chip buses and/or point-to-point interconnections.
Extended task graph
To solve the problem, we extend the input task graph (TG) to an extended
task graph (ETG). Figure 10-2(a) shows a TG of H.263 encoder system and
Figure 10-2(b) shows an ETG extended from the TG. As shown in Figure
10-2(b), we replace an edge of the TG into two communication nodes representing
write and read data transactions. For instance, the edge between Source
and MB_Enc of Figure 10-2(a) is transformed to one communication node for
the write operation, and another communication node for read operation,
in Figure 10-2(b).
3.2. ILP formulation
Before explaining our ILP formulation, we explain our notations used in it.
3.2.1. ILP for dynamic software behavior
The delay of the communication nodes in the ETG can be modeled in the
ILP formulation as follows:
communication time of communication node
of communication data
start time of node
finish time of node
where n is the size