Registering urban point clouds is a quite challenging task due to the large-scale, noise and data incompleteness of LiDAR scanning data. In this paper, we propose SARNet, a novel semantic augmented registration network aimed at achieving efficient registration of urban point clouds at city scale. Different from previous methods that construct correspondences only in the point-level space, our approach fully exploits semantic features as assistance to improve registration accuracy. Specifically, we extract per-point semantic labels with advanced semantic segmentation networks and build a prior semantic part-to-part correspondence. Then we incorporate the semantic information into a learning-based registration pipeline, consisting of three core modules: a semantic-based farthest point sampling module to efficiently filter out outliers and dynamic objects; a semantic-augmented feature extraction module for learning more discriminative point descriptors; a semantic-refined transformation estimation module that utilizes prior semantic matching as a mask to refine point correspondences by reducing false matching for better convergence. We evaluate the proposed SARNet extensively by using real-world data from large regions of urban scenes and comparing it with alternative methods. The code is available at https://github.com/WinterCodeForEverything/SARNet.