Section: New Results

On Using Lower Bounds for Discrimination of Utility/Futility of Loop Fusion

Participants : Samyam Rajbhandari [OSU] , Martin Konk [OSU] , P. Sadayappan [OSU] , Robert J. Harrison [Stonybrook] , Fabrice Rastello.

Fusion is an important loop transformation for data locality enhancement. However, it is very challenging to determine which of a set of possible fusion choices is best. In this paper, we pursue a novel approach to addressing this problem. Instead of the conventional approach of explicitly modeling different possible fused loop configurations and modeling the expected performance with each, we instead use lower bounds modeling to characterize conditions where fusion might have utility and where it will be futile because the maximal possible improvement from fusion is much lower than the minimal data movement overheads for each of the unfused components. We successfully demonstrate the use of such a methodology with two practically important codes from the quantum chemistry domain, i) with the affine 4-index transform code, and ii) unstructured tree operations with the MADNESS framework.

This work is the result of the collaboration  8.1 with OSU.