A Non-Systolic Matrix Product Algorithm
Abstract
Most parallel matrix-matrix product algorithms for MIMD architectures are systolic. These algorithms can be adapted to be used on a general purpose architecture, such as the CHiP or cube machines. When the processors already contain the matrices, the algorithm can still be used by modifying it to circulate the data as if it was being fed in from an external source. We present a non-systolic matrix product algorithm in which the data movement is not the circulation pattern of the adapted systolic algorithms. Rather, it uses technique similar to Strassen's algorithm. The running time is O(n) using n- squared processors for nxn matrices. We compare this algorithm to a systolic algorithm and give experimental results. Keywords: Wavefront array processors; XX Parallel language.
Document Details
- Document Type
- Technical Report
- Publication Date
- Nov 01, 1985
- Accession Number
- ADA196418
Entities
People
- Philip A. Nelson
Organizations
- University of Washington