End-to-End DNN Inference on a Massively Parallel Analog In Memory Computing Architecture