Section: New Results

Parwiz: dynamic data dependence analysis

Participants : Alain Ketterlin, Philippe Clauss.

We have continued working on dynamic data-dependence analysis during this year, especially on increasing the scope of our tool (called Parwiz). For instance, Parwiz is now able to suggest several program transformations (like loop distribution) that enable loop vectorization. It uses an algorithm known as codegen (developed by Allen & Kennedy), but the novelty is that it applies the algorithm to dependence graphs that are built empirically, by running the program on selected input data sets. As far as we know, Parwiz is the first tool able to suggest loop transformations.

We have also developed several other empirical analysis. One of these focuses on loops that are not parallel, but whose iterations present significant parallelism provided the program explicitly schedules the various iterations. This still lacks a suitable cost model to estimate the potential gain, but gives significant insight into the behavior of a given non-parallel loop.

This work has been presented at the MICRO-45 conference held in Vancouver on december 1–5 2012 [18] .