IMPLEMENTATION OF THE BLAS LEVEL-3 AND LINPACK BENCHMARK ON THE AP1000

This paper describes an implementation of Level 3 of the Basic Linear Algebra Subprogram (BLAS-3) library and the LINPACK benchmark on the Fujitsu AP1000. The performance of these applications is regarded as important for distributed memory architectures such as the AP1000. We discuss the techniques involved in optimizing these applications without significantly sacrificing numerical stability. Many of these techniques may also be applied to other numerical applications. They include the use of software pipelining and loop unrolling to optimize scalar processor computation, the utilization of fast communication primitives on the AP1000 (particularly row and column broadcasting using wormhole routing), blocking and partitioning methods, and 'fast' algorithms (using reduced floating point operations). These techniques enable a performance of 85-90 % of the AP1000's theoretical peak speed for the BLAS Level 3 procedures and up to 80 % for the LINPACK benchmark.