Section: New Results


Participant : Tomofumi Yuki.

Loop transformation frameworks using the polyhedral model have gained increased attention since the rise of the multi-core era. We now have several research tools that have demonstrated their power on important kernels found in scientific computations. However, there remains a large gap between the typical kernels used to evaluate these tools and the actual applications used by the scientists.

PolyApps is an effort to collect applications from other domains of science to better establish the link between the compiler tools and “real” applications. The applications are modified to bypass some of the front-end issues of research tools, while keeping the ability to produce the original output. The goal is to assess how the state-of-the-art automatic parallelizers perform on full applications, and to identify new opportunities that only arise in larger pieces of code.

We showed that, with a few enhancements, the current tools will be able to reach and/or exceed the performance of existing parallelizations of the applications. One of the most critical element missing in current tools is the ability to modify the memory mappings.