Static computation of standard linear algebra subroutines for PTAH

The compilation of communication and scheduling is an attractive approach to improve the performance of computers because this avoids some hardware and software levels of interpretation. A key question to face is: can we use such an approach for general purpose applications? We try to answer this question by analysing the relevance of the approach for the Linpack library. Static and dynamic statistics on the delta references made in the Linpack library are presented. We introduce the architecture of PTAH as an example of a massively parallel architecture that requires compiled communication and scheduling. We confront the PTAH architecture with Linpack and we give an example of the implementation of some typical subroutines. Especially, we examine the data layout and the efficiency of this implementation.<
>