high performance computing on graphics processing units: hgpu.org

Computer vision (CV) is widely expected to be the next "Big Thing" in mobile computing. For example, Google has recently announced their project "Tango", a 5-inch Android phone containing highly customized hardware and software designed to track the full 3-dimensional motion of the device as you hold it while simultaneously creating a map of the environment. One of the problems yet to solve is how to transfer demanding state-of-the-art computer vision algorithms -designed to run on powerful desktop computers with several graphics processing units (GPUs) – onto energy-efficient, but slow embedded GPUs found in mobile devices. This project investigates ways of speeding up computer vision kernels and applications through optimisation and parallelisation. We took a representative example of a CV application, the KinectFusion, and we ported it to a mobile platform using OpenCL. Then, we conducted a performance evaluation, identifying performance bottlenecks and further optimise performance. We finally broaden our focus and studied its performance on a different platform to evaluate the performance portability of our optimisations.