Résumé
We measure and analyze the instruction level parallelism that condition the running-time performance of core numerical subroutines. We propose PerPI, a programmer oriented tool to fill the gap between high level algorithm analysis and machine dependent profiling tools and which provides reproducible results.