high performance computing on graphics processing units: hgpu.org

Computational fluid dynamics (CFD) methods have been employed in the studies of subjects such as aeroacoustics, gas dynamics, turbo machinery, viscoelastic fluids, among others. However, the need for accuracy and high performance resulted in methods whose solutions are becoming increasingly more complex. In this context, feature extraction and visualization methods play a key role, making it easier and more intuitive to explore and analyze the simulation data. Feature extraction methods detect and isolate relevant structures in the context of data analysis. In the case of flow analysis, these structures could be pressure isocontours, vortex cores, detachment lines, etc. By assigning visual attributes to these structures, visualization methods allow for a more intuitive analysis through visual inspection. Traditionally, CFD methods represent the solution as piecewise linear basis functions defined over domain elements. However, the evolution of CFD methods has led to solutions represented analytically by higher-order functions. Despite their accuracy and efficiency, data generated by these methods are not compatible with feature extraction and visualization methods targeted to linearly interpolated data. An alternative approach is resampling, which allows the use of existing low order feature extraction and visualization methods. However, resampling is not desirable since it may introduce error due to subsampling and increase memory consumption associated to samples storage. To overcome these limitations, attention has recently been given to methods that handle higher-order data directly. The main contributions of this thesis are two methods developed to operate directly over higher-order data. The first method consists of an isocontouring method. It relies on a hybrid technique that, by splitting the isocontouring workload over image and object space computations, allows for interactive data exploration by dynamically changing isovalues. The second method is a line-type feature extraction method. The search for features is accomplished using adaptive subdivision methods driven by the evaluation of the inclusion form of the parallel vectors operator. Both methods were designed to take advantage of the parallelism of current graphics hardware. The obtained results are presented for synthetic and real simulation higher-order data generated with the discontinuous Galerkin method.