8002824262

12 Włodzimierz Bielecki, Krzysztof Kraska

Mathematica package. The CodeGenerationModule is responsible for the generation of source codę destined for the target architecture; there is the explicit assumption that the tool will be able to generate output codę according to the OpenMP2.0 standard.

5. Conclusion

The estimation of data locality for parallel threads of a loop extracted by means of the method introduced in [1] is a necessary activity to assure optimal performance for a program adjusted to and executed in a targeted architecture. The experiments presented in this paper targeted to an embedded system environment indicate on the identical gains achieved through the increased utilization level of DCU for sources where well-known optimization techniąues to improved data locality were applied.

An important observation obtained from the research is that other software compo-nents of the system simultaneously worked during the program execution have an unfo-reseen impact on the cache utilization. For this reason, it is correctly to assume that the available cache size is different for various configurations of the system. To optimize cache utilization, it is worth to estimate the actual size of available cache memory.

Let be noticed that the limitations of using the results of our work go from the limi-tations of the method introduced in [3]. It applies group-reuses only to separate array references that indexes expressions produce the same coefficient matrix A. Our effort is targeted toward the expansion of the presented method to operate the cases where: A, z, +c, = Aj/j +c₂.