Multipipeline networking for compound vector processing

Journal Article · Fri Jan 01 04:00:00 EST 1988 · IEEE Trans. Comput.; (United States)

DOI:https://doi.org/10.1109/12.75149· OSTI ID:5341011

Hwang, K; Xu, Z

An efficient vector processing technique is proposed based on a new concept of multipipeline networking, which is generalized from the techniques of pipeline chaining and systolization. A pipeline net can be viewed as a programmable systolic array, which can be dynamically reconfigured to evaluate various vector compound functions. In other words, pipeline nets can best match parallelism in user program graphs with the efficient use of multiple functional pipelines in a vector supercomputer. The reconfigurability of a pipeline net provides the flexibility in implementing most Livermore loops, matrix algebra, complex arithmetic, polynomial evaluation, and signal/image processing algorithms. This paper presents the design principles of pipeline nets and provides programming, compiling, and run-time techniques for converting scientific programs into pipeline net implementations. Performance analysis of the pipeline net is provided with projected performance of various Livermore loops implemented with pipeline nets of various sizes.

• Research Organization: Dept. of Electrical Engineering-Systems, Univ. of Southern California, Los Angeles, CA 90089 (USA)
• OSTI ID: 5341011
• Journal Information: IEEE Trans. Comput.; (United States), Journal Name: IEEE Trans. Comput.; (United States) Vol. 37:1; ISSN ITCOB
• Country of Publication: United States
• Language: English

Related Subjects

99 GENERAL AND MISCELLANEOUS
990210* -- Supercomputers-- (1987-1989)
ALGORITHMS
ARRAY PROCESSORS
COMPUTER CODES
COMPUTER NETWORKS
DATA-FLOW PROCESSING
EXECUTIVE CODES
GRAPHS
IMAGE PROCESSING
MATHEMATICAL LOGIC
PARALLEL PROCESSING
PERFORMANCE TESTING
PROCESSING
PROGRAMMING
TESTING
VECTOR PROCESSING