Programmable In-Memory Computing Accelerator for Low-Precision Deep Neural Network Inference