Accelerating an implicit ocean model using CUDA C

In this study, we developed an ocean model named GPU-IOCASM (GPU-Implicit Ocean Current and Storm Surge Model), which employs the finite difference method with implicit iteration to ensure simulation stability. Additionally, it incorporates an online nesting for multi-layer computational grids, allowing localized grid refinement in critical regions to enhance simulation accuracy. To maximize GPU parallelism and minimize memory overhead, we optimized the residual update algorithm, applied a mask-based conditional computation method, and designed an adaptive iteration count prediction strategy. When the simulation reaches a designated output time, relevant variables are copied from GPU memory to host memory, while the GPU proceeds with the next computation without waiting for the I/O operation to complete. This process is designed to run asynchronously in most cases, ensuring that data transfer and CPU-side operations do not interfere with GPU-based computation. Verification results demonstrate that GPU-IOCASM's simulation results exhibit strong agreement with both observed data and SCHISM’s results, confirming its reliability and precision. Furthermore, GPU-IOCASM achieves a remarkable speedup of over 312 times compared with traditional CPU-based approaches. Unlike traditional GPU acceleration methods that require frequent data transfers between the CPU and GPU, GPU-IOCASM is designed to perform as much computation as possible on the GPU, thereby minimizing data transfer overhead and improving computational efficiency.