Section: New Results

The Multifor programming construct

Participants : Philippe Clauss, Imèn Fassi, Yosr Slama, Matthieu Kuhn.

We have proposed a new programming control structure called “multifor”, allowing to take advantage of parallelization models that were not naturally attainable with the polytope model before. In a multifor-loop, several loops whose bodies are run simultaneously can be defined. Respective iteration domains are mapped onto each other according to a run frequency – the grain – and a relative position – the offset –. Execution models like dataflow, stencil computations or MapReduce can be represented onto one referential iteration domain, while still exhibiting traditional polyhedral code analysis and transformation opportunities. Moreover, this construct provides ways to naturally exploit hybrid parallelization models, thus significantly improving parallelization opportunities in the multicore era. Traditional polyhedral software tools are used to generate the corresponding code. Additionally, a promising perspective related to non-linear mapping of iteration spaces has also been developed, yielding to run a loop nest inside any other one by solving the problem of inverting “ranking Ehrhart polynomials”.

This work is the PhD work of Imèn Fassi, who started her work this year and who is co-advised by Yosr Slama, Assistant Professor at the University El Manar in Tunis, Tunisia, and Philippe Clauss. A first publication of this topic has been accepted at the IMPACT workshop that will be held in conjunction with the HIPEAC conference in Berlin, Germany, January 2013.

Figure 3. Red-Black Gauss-Seidel Multifor Iteration Space