Large-field high-resolution electron tomography enables visualizing detailed mechanisms under global structure. As field enlarges, the distortions of reconstruction and processing time become more critical. TxBR has proposed a curvilinear projection model, which can dramatically improve the quality of reconstruction. But its computation is more complex and time-consuming. Furthermore, previous parallel strategies are not suitable for curvilinear projection model. In this work, a block iterative parallel algorithm using curvilinear projection model on GPU platform was proposed. By studying the locality of curvilinear projection model, we proposed a vertical data decomposition method. We also adopt a page-based data transfer scheme to reduce the processing time. Experimental results show that our method can yield speedups of approximate 40 times.