Design of a Memory Framework for Neural Network Accelerators that Utilizes Their Calculable Memory Access Patterns