Graph Neural Networks (GNNs) are powerful to learn representation of graph-structured data, which fuse both attributive and topological information. Prior researches have investigated the expressive power of GNNs by comparing it with Weisfeiler-Lehman algorithm. In spite of having achieved promising performance for the isomorphism test, existing methods assume overly restrictive requirement, which might hinder the performance on other graph-level tasks, e.g., graph classification and graph regression. In this paper, we argue the rationality of adaptively emphasizing important information. We propose a novel global attention module from two levels: channel level and node level. Specifically, we exploit second-order channel correlation to extract more discriminative representations. We validate the effectiveness of the proposed approach through extensive experiments on eight benchmark datasets. The proposed method performs better than the other state-of-the-art methods in graph classification and graph regression tasks. Notably, It achieves 2.7% improvement on DD dataset for graph classification and 7.1% absolute improvement on ZINC dataset for graph regression.