Optimization through recomputation in the polyhedral model

Many modern (mobile) systems involve memory intensive computations. External memory accesses are costly when it comes to the execution time and energy consumption of a program. To overcome this, we usually apply tiling to improve data locality and data reuse in internal memories. In the research reported in this paper we add the possibility to recompute data rather than storing temporary results, and demonstrate that this can have a positive e ect on the overall application performance. To achieve this we represented recomputation in the Polyhedral model by extending Polly. We experimentally veri ed the e ectiveness of recomputation on a pair of Convolutional Neural Network layers, when applying loop tiling, loop fusion, and recompute