This project focuses on utilizing parallel architectures within graphical processing units (GPUs) to fill the need for real-time interaction with large datasets in Computer-Aided Design (CAD) systems. GPU-accelerated multi-display applications for visualizing large 3D datasets are important because new scientific discoveries and dynamic investigations of complex digital contents rely on visualization results, favoring with latencies measured in milliseconds rather than seconds and the rich information displayed with high-resolution on multiple screens. The potential of this work is an end-to-end solution for visualizing large 3D datasets on a commodity desktop and providing high-speed interactions and visual explorations of the datasets while having the advantage on cost savings.
Even more specifically, the algorithms and visualization systems developed through this research will allow engineers and designers to interactively create, change, and visually analyze large datasets. By providing a parallel computational foundation on visualization further innovation in CAD research is possible. Graduate students involved will be trained in the principles and uses of advanced computer by designing and implementing GPU algorithms. In addition, creating applications that make good use of GPUs also impacts interactive learning, especially on graphics-assisted and multi-dimensional educational applications, by providing breakthrough technology to increase the performance of educational applications while displaying complex graphical contents.
Traditionally, acceleration algorithms for rendering (e.g., view-dependent mesh simplification) are implemented sequentially in visualization applications, which are not suitable for fast processing the model, composed of large amount of geometric primitives. The algorithms developed in this project will support primitive-level parallelization on GPUs and enable dynamic adjustments of detail levels of 3D meshes upon the changes of camera configurations. Though multi-GPU applications are available commercially, load-balancing issues among GPUs have not been successfully addressed. Imbalanced workload distributions hurt overall performance and affect the quality of mesh renderings. This research will develop a screen-partitioning balancer to balance GPU workloads and employ a pipelined approach to conduct GPU-to-GPU and CPU-to-GPU streaming for efficient data management. The efficiencies of inter-device communications will be given an in-depth evaluation.
