Predictable Dynamic Behaviour in NoC-based Multiprocessor Systems-on-Chip
Modern streaming multimedia applications are becoming increasingly complex and dynamic.
Still, designers want to give guarantees about the quality and performance of their applications. A combination of model-based design methods and predictable hardware architectures
enables designers to develop individual application components (jobs) in a way that it is
guaranteed that real-time constraints are satisfied, even when multiple jobs are concurrently
running on a single hardware platform.
A multiprocessor Sytem-on-Chip is considered, which consists of a heterogeneous set of
processing elements that are interconnected through a Network-on-Chip. Existing design
methods are often based on the Synchronous Dataflow (SDF) model of computation, which
offers the necessary design-time predictability. However, SDF can not adequately model jobs
that contain dynamic behaviour, like conditionals and loops with data-dependent bounds.
By making dynamic behaviour explicit in the specification model of a job, it is possible to
estimate the needed amount of platform resources more accurately and thus arrive at more
efficient job implementations.
This thesis proposes an extension to SDF, called Predictable Dynamic Dataflow (PDDF).
In PDDF it is p ossible to incorp orate dynamic b ehaviour in a job's specification by the use
of special constructions, while it is still possible to guarantee bounds on timing and memory
usage. PDDF introduces two dynamic constructs: the conditional and the data-dependent
iteration. Both constructs behave like SDF towards their environment, so they can be plugged
into any existing SDF graph. Moreover, they are fully modular, in the sense that they can be
nested in any possible way. The semantics of these constructs are explained and it is shown
how the use of PDDF can lead to more efficient implementations compared to the SDF-based
approach. To illustrate this, an implementation of an H.263 video decoder is used as a test