Using Programmable Graphics Hardware to implement the Fuzzy C-Means Algorithm

The thesis of this work is that programmable graphics hardware can be used to implement complex iterative algorithms. This is achieved by manipulating the graphics processing pipeline, which was design specifically for the graphical rendering process, to perform more general purpose applications. In doing so, the parallelism of the graphics hardware can be to achieve performance gains compared to the traditional central processing unit. To explore this, the NVIDIA GeForceFX programmable graphics pipeline was used to implement the fuzzy cmeans clustering algorithm. The fuzzy c-means clustering algorithm is used in pattern recognition applications, such as image segmentation. However, the computation time of the algorithm becomes too large for complex data sets. Because the algorithm contains several steps in which identical calculations are performed on a large number of vectors, it is a suitable candidate for parallel processing on the graphics hardware to achieve a more time efficient computation. This work will show the use of the graphics pipeline to implement the fuzzy c-means algorithm. A stream programming model is developed that utilises the fragment processor and texture memory of the programmable graphics hardware to realise this goal. The texture memory is used to store data in a manner similar to arrays used in programs executed on the central processing unit. The fragment processor is used to process the data in parallel. This work then explores the use of the high-level stream programming package BrookGPU to implement the stream model. Subsequent testing obtained timing results in which the BrookGPU implementation took approximately eight times as long to run as a comparable CPU implementation. This was found to be due to unnecessary copying of data between the GPU and CPU. For this reason an different implementation was then pursued. An OpenGL implementation is then explored, which was found to have the flexibility required for an efficient algorithm. This flexibility was gained from the ability to use the underlying OpenGL functions directly. The timing analysis was that was conducted found that the computation time was reduced by approximately fifty percent compared to an equivalent CPU implementation. This work has found that it is not only possible to implement the fuzzy cmeans algorithm on programmable graphics hardware, but that doing so has the