Accelerating Unstructured Grid-Based Seismic Modeling on GPU

Seismic modeling is more accurate and flexible for complex geological model based on unstructured grid, which is tessellated by simple shapes in an irregular pattern. Owe to the irregularity, there are mainly two challenges to accelerate unstructured grid-based numerical methods on graphics processing units (GPUs): race condition and global memory coalescence. We propose a scheme composed of three parts to tackle these two problems: firstly apply centroidal Voronoi tessellation (CVT) method to optimize the grid, relaxing the irregularity, secondly develop a novel coloring method to avoid race conditions, thirdly develop a renumbering method of cells and vertexes to increase the global memory coalescence. We use the acoustic grid method to demonstrate the efficiency of our scheme, which also applies to other numerical methods based on unstructured grid.