Software/Hardware Tradeoffs in the Speedup of Color Image Processing Algorithms

Data parallel image processing algorithms have numerous uses in many real time applications. Depending on the complexity of the computations involved, these algorithms may take considerable amounts of time to complete. Since the algorithms are performed in real time, the end user is negatively impacted by the extended execution times. Fortunately, there are many different ways available in hardware and software to improve the speed of these algorithms. This thesis looks at several different methods of improving the speedup of color image processing algorithms and compares the tradeoffs among them. The methods for increasing the execution time of an algorithm include implementing Single Input Multiple Data (SIMD) instructions, using Posix threads to code across several processors, and using a stream based multichannel framework to implement the algorithms on an FPGA. Each of the above methods had advantages and disadvantages, yet all approaches were found to introduce a significant speedup over the single core baseline tests. These methods were completed on a number of different images to examine the effects of workload on the efficiency of the implementations. The application of these speedup techniques yielded excellent results leading to speedups of greater than 3.85 times in software and 5.8 times in hardware. In each of the software tests, the output image had a 2-d correlation coefficient (CORR2) of 1.0000. When implementing the algorithms in hardware using implementation specific approximations, the correlation coefficient of the output image was still an acceptable 0.99 or higher