Static Scheduling of Weight Programming for DNN Acceleration with Resource Constrained PIM

Most existing architectural studies on ReRAM-based processing-in-memory (PIM) DNN accelerators assume that all weights of the can be mapped to crossbar at once. However, these are over-idealized. ReRAM resources for calculation limited because technological limitations, so multiple weight mapping procedures required during inference process. In this paper, we propose a static scheduling framework which generates between and cells with minimum runtime programming cost. We first build latency model by simultaneously considering patterns, operations, PIM architecture characteristics. Then, is used in searching process obtain an optimized weight-to-OU table online latency. Finally, OU scheduler coordinate activation sequences OUs crossbars perform computation correctly. Evaluation results show proposed significantly reduces overhead overall various models different input data sets.