Convolutional Neural Networks (CNNs) have been widely adopted in raster-based urban flow analytics by virtue of their capability in capturing nearby spatial context. By revisiting CNN-based methods for different analytics tasks, we expose two common critical drawbacks in the existing uses: 1) inefficiency in learning global context, and 2) overlooking latent region functions. To tackle these challenges, in this paper we present a novel framework entitled DeepLGR that can be easily generalized to address various urban flow analytics problems. This framework consists of three major parts: 1) a local context module to learn local representations of each region; 2) a global context module to extract global contextual priors and upsample them to generate the global features; and 3) a region-specific predictor based on tensor decomposition to provide customized predictions for each region, which is very parameter-efficient compared to previous methods. Extensive experiments on two typical urban analytics tasks demonstrate the effectiveness, stability, and generality of our framework.