Optimal data allocation algorithm for loop-centric applications on scratch-PAD memories

The traditional way of managing SPM with arrays is fetching data tiles that are required for the current computation from the off-chip memory. Yet, this method is not the best solution for the array intensive applications. In this paper, we propose a dynamic programming algorithm that can optimally minimize the data transferring cost and the memory access cost for a regular loop. According to the experimental results, the DLAA algorithm can reduce time and energy of the total memory access cost for loops by an average 35-40% compared with the traditional data tiling algorithms.