high performance computing on graphics processing units: hgpu.org

Direct manipulation using pointing devices commonly relies on basic interaction tasks such as selection and precise cursor positioning. Traditionally, such tasks use geometric attributes computed on the basis of application-dependent intersection algorithms on the CPU. However, with the recent ability of using GPUs to perform geometry modeling tasks and even to create new primitives on-the-fly, geometric attributes computed on the CPU are often invalid. We analyze a new paradigm for correctly computing such attributes based on providing, for each pixel of the rendered models, application-defined data and elements of discrete differential geometry computed solely on the GPU. We validate this by showing how many direct manipulation tasks presented in the literature require only these data, and present an interaction framework that exploits such paradigm. Implementation results are presented.