The rapid emergence of large-scale atlas-level single-cell RNA-seq datasets presents remarkable opportunities for broad and deep biological investigations through integrative analyses. However, harmonizing such datasets requires integration approaches to be not only computationally scalable, but also capable of preserving a wide range of fine-grained cell populations. We have created Portal, a unified framework of adversarial domain translation to learn harmonized representations of datasets. When compared to other state-of-the-art methods, Portal achieves better performance for preserving biological variation during integration, while achieving the integration of millions of cells, in minutes, with low memory consumption. We show that Portal is widely applicable to integrating datasets across different samples, platforms and data types. We also apply Portal to the integration of cross-species datasets with limited shared information among them, elucidating biological insights into the similarities and divergences in the spermatogenesis process among mouse, macaque and human.