Design of the Scientific Subroutine Library for the Fujitsu VPP300

A research agreement between the Australian National University and Fujitsu Japan has led to the development of a library of parallel mathematical subroutines and the extension of the library of single processor routines for the Fujitsu VPP300. The Fujitsu VPP300 provides a very sophisticated architecture combining vector processors in parallel by a crossbar switch. Very high performance can be obtained if carefully designed algorithms are used which exploit the full performance of the hardware. New algorithms have been developed and performance results are provided for eigenvalue problems, linear solvers and fast Fourier and wavelet transforms. 1. Introduction With 13 vector processors, the Fujitsu VPP300 at the Australian National University has a peak spead of 29 Gflop/s and a main memory of 14 Gbytes. Thus the VPP300 is ideally suited to processing very large and demanding simulation and data analysis tasks. Each processor has a 7 ns cycle time and contains ffl A scalar unit (SU)..