The impact of computer architecture features on image processing application execution times: A case study using MPEG image sequence compression on the IBM SP2.

Many image-processing applications require special-purpose hardware to run in real time. Others can run on general-purpose computers only on small images. The use of image-processing programs becomes more common as general-purpose computers advance so they run these programs in real time on larger images. The field of mathematics called image algebra defines a complete set of data and operations for expressing any image-to-image transformation. This mathematical foundation has previously been used to guide the development of special-purpose image processors. The research reported in this thesis expands on the previous work to show how to apply this theoretical and practical background to identify specific features that could improve the performance of general-purpose computers when executing a wide variety of image-processing programs. We used an MPEG2 encoding program as a benchmark to test the validity of the approach with experiments performed on an IBM SP2 scaleable parallel computer. The nodes of the SP2 used for the experiments include the Power2 processor which runs the RS6000 instruction set. The theoretical background suggested, and the experiments verified, specific enhancements to decrease the execution time for the fundamental image-algebra operations. The most significant impact would be to increase the number of moderate precision (8--16 bits) integer arithmetic operations that can be executed each clock cycle. This can be accomplished with additional fixed-point units, instructions to convert from fixed-point to floating-point format, or floating-point units that can operate on fixed-point data. Compound instructions for integer multiply-accumulate, add-max, and multiply-max are also recommended to better support the fundamental image-algebra operations. The theoretical background was also applied to identify and evaluate similar improvements to sub-word parallel instruction-set extensions such as those contained in MMX, VIS, Altivec, and MAX-2. In addition, enhancements to the communications structure were identified to improve the ability of message-passing distributed-memory MIMD computers to execute parallelized image-processing programs. We concluded that the image algebra provides a solid theoretical foundation that suggests effective ways to improve general-purpose computers so they execute image-processing programs faster.PhDApplied SciencesComputer scienceElectrical engineeringSystems scienceUniversity of Michigan, Horace H. Rackham School of Graduate Studieshttp://deepblue.lib.umich.edu/bitstream/2027.42/132463/2/9963887.pd