Benchmark framework of compute-in-memory based accelerators for deep neural network (on-chip training chip focused)