In this paper, we propose a strategy that can simultaneously reduce the amount of AD and DA conversions, while also allowing for a reduction on the number of total ReRAMs required to compute a NN. We achieve this by deploying a crossbar-friendly pruning technique, and show how one can reprogram the ReRAMs to compute the activation functions and pooling layers. Experiments on real-world Human Activity Recognition and speech recognition datasets demonstrate that our device outperforms analog and off-the-shelf digital approaches by up to 17.8x, allowing for flexibility with reduced power and high performance.