Automatic benchmarking and optimization of codes: an experience with numerical kernels

Abstract New algorithms are constantly devel-oped in search of better or faster results. Many variants of code are often tried while searching for the best solution. When the number of code vari-ants or possible input parameters is very high, the process of benchmarking and analysis of results can become cumbersome and error prone. We present a tool for automatic benchmarking which manages a database of possible parameters and the results ob-tained for them. This tool can automatically choose the optimum code and create a target library. Us-ing it we have generated a library specialized in the operation on very small matrices which is useful in multimedia codes