Inria / Raweb 2004
Project-Team: DaRT

Search in Activity Report, year 2004:
HELP

INDEX

Project-Team : dart

Section: Scientific Foundations


Keywords: Scheduling, Mapping, Compilation, Optimization, Heuristics, Power Consumption, Dataparallelism.

Optimization Techniques

Participants: Pierre Boulet, Jean-Luc Dekeyser, Philippe Dumont, Philippe Marquet, Ashish Meena, Smaïl Niar.

We study optimization techniques to produce a schedule and a mapping of a given application onto a hardware SoC architecture. These heuristic techniques aim at fullfilling the requirements of the application, whether they be real time, memory usage or power consumption constraints. These techniques are thus multi-objective and target heterogeneous architectures.

We aim at taking advantage of the parallelism (both data-parallelism and task parallelism) expressed in the application models in order to build efficient heuristics. Our application model has some good properties that can be exploited by the compiler: it expresses all the potential parallelism of the application, it is an expression of data dependences –so no dependence analyzis is needed–, it is in a single assignment form and unifies the temporal and spatial dimensions of the arrays. This gives to the optimizing compiler all the information it needs and in a readily usable form. Many optimization techniques have been studied that can be useful in our case. These techniques cover several fields of compiler construction:

Contributions

We focus on two particular subjects in the optimization field: dataparallelism efficient utilization and multi-objective hierarchical heuristics.

Dataparallel Code Transformations

In some of our previous works have studied Array-OL to Array-OL code transformations [32][67][40][39]. Array-OL [37][38] is a dataparallel language dedicated to systematic signal processing. It allows a powerful expression of the data access patterns in such applications and a complete parallelism expression. It is at the root of our model of applications.

The code transformations that have been proposed are related to loop fusion, loop distribution or tiling but they take into account the particularities of the application domain such as the presence of modulo operators to deal with cyclic frequency domains or cyclic space dimensions (as hydrophones around a submarine for example).

We currently study the relations of the Array-OL model with other computation models such as Kahn Process Networks [50][51] and multidimensional synchronous dataflow [57][56].

We pursue the study of such transformations with three objectives:

This works is the subject of Philippe Dumont's Ph. D. Thesis.

Multi-objective Hierarchical Scheduling Heuristics

When dealing with complex heterogeneous hardware architectures, the scheduling heuristics usually take a task dependence graph as input. It is the case in the AAA methodology [66][65][45] that is implemented in the SynDEx [64] tool. Both our application and hardware architecture models are hierarchical and allow repetitive expressions. We believe that we can take advantage of these hierarchical and repetitive expressions to build more efficient schedules. We call this approach globally irregular, locally regular (GILR). We have shown in [33] that GILR heuristics can improve the optimization in several ways:

Further more, local optimizations (contained inside a hierarchical level) will surely decrease the communication overhead and allow a more efficient usage of the memory hierarchy. We aim at integrating the dataparallel code transformations presented before in a global heuristic in order to deal efficiently with the dataparallelism of the application by using repetitive parts of the hardware architecture.

Furthermore, in embedded systems, minimizing the latency of the application is usually not the good objective function. Indeed, one must reach some real time constraints but it is not useful to run faster than these constraints. It would be more interesting to improve the resource usage to decrease the power consumption or the cost of the hardware architecture. We will thus study multi-objective techniques to build schedules that respect the real time constraints of the application while minimizing the resource usage.

Ashish Meena is working towards a Ph. D. on this subject. Smaïl Niar, associate member of the project from the university of Valenciennes, is studying various techniques to reduce power consumption in embedded systems. This research covers:

We plan to use these results to build our scheduling heuristic.


previous
next