high performance computing on graphics processing units: hgpu.org

Finite element and finite volume methods on unstructured meshes offer a powerful approach to solving partial differential equations in complex domains. It has diverse application in areas such as industrial and geophysical fluid dynamics, structural mechanics, and radiative transfer. A key strength of the approach is the unstructured meshes exibility in conforming to complex geometry and to smoothly vary resolution throughout the domain. Adaptive mesh methods further enhance this capability by allowing the mesh to be locally modified in response to local estimates of simulation error. The ability to locally control simulation error plays an important role in both optimally exploiting available computational resources to achieve the most accurate solution feasible, or simulating a process to within design/safety guidelines for example. This report focus on the anisotropic adaptive mesh operations of coarsening and refinement on meshes of 2D simplexes. The novelty of this work centres on recasting the coarsening and refinement algorithms, which were developed for serial execution on CPU’s, into a form better suited to the massively parallel nature of GPU’s. An impressive speedup has been achieved when compared against the best known multi-threaded CPU implementation on top of the range hardware.