Currently, we focus on the implementation of new techniques for a source to source compiler that enable the automatic conversion of programs that operate on so-called dense matrice into equivalent program operating on so-called sparse matrices. The latter are matrices in which the presence of many zeros can be exploited to reduce storage requirements and computational time. Clearly, more powerful transformations than the traditional program transformations are required, because the data structures must be addapted in order to exploit the characteristics of data.
The main ideas have been elaborated and are proposed on several congresses. A mechanical conversion is certainly possible. However, supporting transformations are required to obtain satisfactory performance. Also `sparsity specific issues' such as fill-in reducing methods and garbage collection must be dealt with.