Comparing the behaviour of basic linear algebra routines on multicore platforms

Abstract

The use of an OpenMP compiler optimized for a multicore system could contribute to obtain programs with satisfactory execution time, but it is possible to have access in a system to more than one compiler, and different compilers optimize different parts of the code at different levels. In this paper, the influence of the compiler used on the performance of linear algebra routines is studied. From the results of the experiments carried out, we conclude a poly-compiling approach is necessary to decide automatically the best compiler.