Smoothed Particle Hydrodynamics (SPH) performs excellently in simulating pollutant transport. However, it faces challenges such as redundant computations and inefficient neighbor searches, particularly in large-scale simulations. To address these issues, we propose a Background Grid-based Particle Dormancy strategy (BGP), which dynamically partitions the computational domain and employs bitwise operations to identify inactive regions, allowing particles in these areas to enter a dormant state, thereby significantly reducing computational costs. Furthermore, to overcome the bottleneck of neighbor search in GPU parallel computing, a GPU-based Multi-threaded Search Strategy for SPH (GMSS) is introduced, extending parallelization from the particle-level computation to the neighbor search process, thereby enhancing computational efficiency. Based on these strategies, a CUDA-based SPH parallel computing framework was developed to optimize large-scale pollutant transport simulations. Experimental results show that the proposed optimization strategies achieved up to a 7x speedup in simulations with millions of particles while maintaining high computational accuracy. This framework provides a robust and efficient solution for large-scale fluid transport simulations.