GPU Usage for Parallel Functions and Contacts in Modelica
Abstract: This thesis investigates two ways of incorporating GPUs in Modelica. The first by automatically generating GPU code for Modelica functions, and the second by using GPU accelerated external code for a contact handling package. Automatic parallelization of functions is desired, as it can potentially accelerate large simulations significantly. Special patterns of nested for-loops in Modelica code are recognized and translated into CUDA kernel functions. Inline integration allows a broader spectrum of models to take advantage of the parallelization, by reducing CPU-GPU transfers. The prototype has been tested and achieved a speed-up factor of up to five compared to the CPU. The contact handling package is capable of handling both complex contact behavior between arbitrarily shaped bodies and large DEM-like simulations, something which Modelica is currently lacking. Attempts to accelerate the package with GPUs were made, with partial success for the broad phase. The package uses Morton encoding for the broad phase, and the narrow phase is based on CSG intersection with BSP trees. Contact response is calculated using a volume dependent method, taking friction, damping and multiple contact points into account. The capability of the package was demonstrated by the fact that both complex contact behavior such as the inversion of the Tippe Top toy and tens of thousands of colliding spheres could be simulated.
AT THIS PAGE YOU CAN DOWNLOAD THE WHOLE ESSAY. (follow the link to the next page)