With the development of computer technology, grid division technology is becoming more mature. Considering the frequent occurrence of floods due to climate change, the broad extents of calculation domains, the wide range of actual terrain, and the study area usually has narrow and long gullies and wide flooding areas, this paper proposes a structured non-uniform grid model with hierarchical topological relationships combined with a high-resolution model based on GPU acceleration to simulate the surface water flow process. High-quality grids affect the calculation accuracy and efficiency of the model. The principle of grid division is designed based on the gradient change of terrain elevation, and key terrain features are detected in the computational domain that requires high-resolution grids to reliably solve shallow water equations. Moreover, local area grids can be statically encrypted, so that the sensitive area of the water level calculation can be captured more accurately, while reducing the number of calculation grids and reducing the calculation cost. The numerical model adopts Godunov-type finite volume method for spatial discretization, uses the second-order TVD-MUSCL format to improve the temporal and spatial accuracy of the model, and uses GPU parallel technology to greatly increase the running speed of the model without reducing the calculation accuracy. The performance of high-resolution models on non-uniform grids is demonstrated by the more accurate simulation of flood inundation time and inundation area through ideal and practical cases. The results show that the numerical model based on the non-uniform grid has good stability, compared with the uniform grid, its running speed is about 2-3 times under the premise of ensuring the simulation accuracy and the efficiency is further improved on the basis of GPU acceleration. The new model is suitable for simulating large-scale flood evolution and urban inundation processes in complex areas, which has good potential in actual large-scale flood simulation.