Design Considerations for Efficient Deep Neural Networks on Processing-in-Memory Accelerators