Abstract: Large-scale seismic simulations are critical for accurate ground motion characterization and real-time hazard assessment. However, such simulations pose significant computational challenges, particularly in regions with complex topography. The curvilinear grid finite-difference method (CGFDM) provides an effective approach for modeling wave propagation in such irregular geometries. To enable CGFDM to fully exploit the massive parallelism of modern heterogeneous architectures, we develop SW-CGFDM3D, a novel computational scheme designed for the Sunway platform using a hardware-algorithm co-design strategy. The proposed method incorporates a multi-level parallelization strategy and a grid-stencil-to-core mapping scheme tailored to the CGFDM. This design enables utilization of the master-slave core groups of the Sunway architecture. The proposed method is validated against analytical solutions and a reference solver in both layered and complex topographic models, showing excellent agreement. In terms of computational performance, the algorithm achieves a significant 17-20× speedup compared to a baseline implementation using only the master cores. To demonstrate its practical utility in computational seismology, we present a real-world simulation of the 2021 Maduo earthquake, incorporating 3D heterogeneous structure, rugged topography, and a finite-fault source. The results confirm the method's accuracy, efficiency, and capability for high-fidelity, rapid seismic hazard assessment. Given its high performance, SW-CGFDM3D serves as an enabling tool for large-scale seismic simulations and timely disaster mitigation.