high performance computing on graphics processing units: hgpu.org

State-of-the-art video analysis systems feature multiple complex processing steps and operate on high resolution images. Intensive computation power is needed for real-time execution. In this project an image change detection application is mapped to a heterogeneous multicore CPU/GPU platform. It is investigated what hardware configuration is required to execute the application in real-time. For optimal execution, i.e. minimum execution time, a choice has to be made which parts of the application to execute on the CPU and which on the GPU. The difference between CPU and GPU hardware architecture styles decreases. Historically CPUs were designed for low latency and GPUs for high throughput. In CPU architectures the trend has clearly turned towards multicore execution. On the other hand, GPU architectures have been optimized for general purpose execution. As a result, it becomes less obvious what processor type best suits an application. Image processing algorithms mostly operate on large data sets, e.g. a full-size image or a set of image features. Therefore most algorithms are very well suited to be implemented on a GPU, which is designed for highly data-parallel applications. Many research project have shown that a significant speedup is possible for various image processing algorithms [PXW10], [CVG08], [ZCW10]. But not every algorithm is equally suited for GPU implementation, since the nature of the algorithm and the GPU architecture are uncooperative.