CPU ACCELERATION AND PERFORMANCE EVALUATION OF IMAGE PROCESSING ALGORITHMS ON CUDA

Abstract

The demand of the huge image data with complex arithmetical operations to be processed in limited time and the developing computational capability of graphic processor, has lead the world to do parallel processing in GPU. In this thesis implementation of image processing algorithms on a massively parallel manner using NVIDIA CUDA is done. Furthermore the analysis of resulting performance gains against the CPU implementation is presented. image processing algorithm often quite complex and require huge arithmetical calculations. NVIDIA has evolved a technology called CUDA "compute unified device architecture" is suitable for program graphic card for general purpose. CUDA is the extension of C language with an added functionality to make program run on graphic cards, using thousands of concurrent threads. The obvious and potential performance improvement using CUDA for image processing algorithm will be investigated here in detail. Additionally optimization strategies are discussed in this thesis. In this dissertation implementation of three algorithm with CUDA, namely Histogram equalization to enhance contrast between fine detail of image, a template matching algorithm to find the similarity between given image knows as template and the search image, if present anywhere and the canny edge detection algorithm used to find the edges in the image is implemented on CUDA to leverage the advantage of GPU massive parallel architecture. -' Furthermore the performance is analyzed by comparing result with their CI'U implementation. The final result shows that it is more effective to use graphic processor to do parallel computing as it perform fast in data parallelism and also more energy efficient