ABSTRACT
A significant amount of execution time in complicated linear algebraic programs is known to be spent in a few low-level operations. Consequently, reducing the overall execution time of an application program leads to the problem of optimizing these low-level operations. Such low-level optimization are highly machine-dependent and are a matter for specialists. A separation has, therefore, been made by the computational linear algebra community: On the one hand highly efficient, machine-dependent building blocks of linear algebra, the BLAS, (basic linear algebra subprograms) are provided on a given computational platform. On the other hand, linear algebra programs attempt to make the most use of those computational blocks in order to have as good performance as possible across a wide range of platforms.
