More Definite Results from the PLuTo Scheduling Algorithm

The PLuTo scheduling algorithm is a well-known algorithm for automatic scheduling in the polyhedral compilation model. It seeks linearly independent affine partition mappings for each statement of a Static Control Program (SCoP), such that total communication is minimized. In this paper we show that this algorithm can be sensitive to the layout of the global constraint matrix thus resulting to varying performances of our target code simply by changing the order of the constraint coefficients. We propose a simple technique that automatically determines the right layout for the constraints and as a result immunizes the PLuTo scheduling algorithm from constraint ordering variations.