high performance computing on graphics processing units: hgpu.org

In the past decade liquid crystal displays (LCD) have taken over the television (TV) and monitor market from cathode ray tube (CRT) display. Compared to CRT displays, LCD offers larger screen sizes, higher resolution, thinner, lighter, and more energy efficient. However, with respect to image quality, LCD does not catch up to CRT display in terms of color accuracy, color range, and dynamic range. Due to the digital nature of LCD, many techniques have been developed from the digital image processing point of view to improve the image quality of LCD. These techniques usually involve algorithms containing pixel-based operations, digital filters, or block-matching, etc. Consequently intensive computations are needed. While dedicated integrated circuits (IC) are used to efficiently realize a proven algorithm for real-time processing on commercial products, it is not applicable for research and development due to lack of flexibility. Realizing a to be analyzed algorithm in high level languages such as MATLAB and executing it on a computer central processing unit (CPU) are a common approach in the algorithm design, prototyping, and evaluation phase. This approach provides complete freedom of tuning algorithms, but on the other hand, processing is usually slow and far from real time. Due to insufficient processing power of CPUs, executing an algorithm takes at least seconds and may take up to tens of minutes for complex algorithms, which makes it time intensive to analyze temporal effects of the algorithm in question. Recent advances on computer graphics processing units (GPU), especially the NVIDIA Compute Unified Device Architecture (CUDA), provide a new approach to solve the flexibility vs. real-time processing dilemma. Current NVIDIA CUDA GPUs can have hundreds of processing units that running in parallel at high execution speed and can be programmed using ‘C for CUDA’, which is an NVIDAextension of the C programming language. In this work we take benefit of such architecture, proposing a platform that allows flexible design and real-time execution of advanced LCD processing algorithms.