Combining Experimental Search and Performance Model for Adaptive Optimization

Abstract. The increasing complexity of hardware features in modern processors makes compilation for high performance very challenging. Per-formance models used by compilers are too simple to take into account this complexity and choose accordingly the most effective optimization sequence. Adaptive compilation is now a widespread approach relying on an explo-ration for optimization sequences or compiler flags and on code execution in order to evaluate precisely performance. The main drawback of this approach is its very high cost, that is partially addressed by efficient search techniques based on genetic or machine learning algorithms. This paper presents a novel approach for adaptive compilation, relying on performance evaluation of only fragments of the code, named constant performance codelets, and on a simple performance model. The search for transformations leading to these codelets is user-defined through prag-mas. We show on three large applications (two numerical simulations and a genomic application) that the performance prediction for the op-timized function is quite accurate and that substantial speed-up can be reached on Itanium2 architecture.