A generalized static data flow clustering algorithm for mpsoc scheduling of multimedia applications

Abstract

In this paper, we propose a generalized clustering approach for static data flow subgraphs mapped onto individual processors in Multi-Processor System on Chips (MPSoCs). The goal of clustering is to replace the static data flow subgraph by a single dynamic data flow actor such that the global performance in terms of latency and throughput is optimized. Through our proposed clustering approach, the scheduling of connected static data flow subgraphs can be coordinated with enclosing system representations in a way that systematically exploits the predictability and efficiency of the static data flow model. Thus, the advantages of static data flow subsystems can be exploited in the context of overall system representations that are based on more general models of computation. At the same time, our approach goes significantly beyond previous approaches to synchronous data flow clustering by providing a quasi-static -as opposed to purely-static -scheduling interface between clustered subgraphs and the enclosing systems. This greatly enhances the power of our techniques in terms of avoiding deadlock, increasing the design space for clustering, and providing for integration with more general models of computation. We show benefits of up to 95% performance improvement for real world examples.

Compositionality in Synchronous Data Flow: Modular Code Generation from Hierarchical SDF Graphs

“…We illustrate this by example. Note that in [4] it is also observed that CSDF is not a sufficient abstraction of composite SDF models, however, the example used in that paper embeds a composite SDF graph into a dynamic data flow model. Therefore the overall model is not strictly SDF.…”

“…This model captures the SDF model given in Figure 3 of [4]. Actor A2 is a composite actor designed so as to consume 2 tokens on each of its input ports and produce 2 tokens on each of its output ports every time it fires.…”

“…More recently, [4] also identify the problem of non-compositionality and propose Cluster Finite State Machines (CFSMs) as a representation of composite SDF. They show how to compute a CFSM for a composite SDF actor that contains standard, atomic, SDF sub-actors, however, they do not show how a CFSM can be computed when the sub-actors are themselves represented as CFSMs.…”

“…Finally, we should emphasize our DAG clustering algorithms solve a different problem than the clustering methods used in [9,4] and other works in the SDF scheduling literature. Our clustering algorithms operate on plain DAGs, as do the clustering algorithms originally developed for SBDs [8,7].…”

“…Our clustering algorithms operate on plain DAGs, as do the clustering algorithms originally developed for SBDs [8,7]. On the other hand, clustering in [4,9] is done directly at the SDF level, by grouping SDF actors and replacing them by a single SDF actor (e.g., see Figure 4 in [4]). This, in our terms, corresponds to monolithic clustering, which is not compositional.…”