Towards more hardware-friendly deep learning