Nineteen ways of systolic matrix multiplication (Q4392896)

From MaRDI portal

!

This is the item page for this Wikibase entity, intended for internal use and editing purposes. Please use the normal view instead:

scientific article; zbMATH DE number 1160806
Language Label Description Also known as
default for all languages
No label defined
    English
    Nineteen ways of systolic matrix multiplication
    scientific article; zbMATH DE number 1160806

      Statements

      Nineteen ways of systolic matrix multiplication (English)
      0 references
      0 references
      0 references
      20 December 1998
      0 references
      matrix multiplication
      0 references
      regular iterative algorithm (RIA)
      0 references
      systolic array synthesis
      0 references
      systolic array performance measures
      0 references
      The authors investigate different ways of systolic matrix multiplication. The systematic systolic synthesis approach by using regular iterative algorithm (RIA) representation is adopted to design nineteen different systolic arrays for matrix multiplication. The RIA is defined and steps of the systolic array design methodology are described: choosing processor and iteration spaces, obtaining a schedule of computation, mapping the points on the index space into the processor space, determining the interconnections between the processors. Various properties of interest can be measured for each processor array obtained through the above procedure. The procedure listed above is used to design the systolic arrays for the matrix multiplication algorithm. Different choices of the iteration space lead to the different types of the systolic arrays. The fifteen different systolic arrays for dense matrix multiplication and the other four arrays for band matrix multiplication are discussed. There are groups of the orthogonal, rectangular and hexagonal arrays. Finally, the different performance measures for the systolic array are discussed (array size, computation time, pipelining and block pipelining period, speedup, efficiency, I/O channels and area-time complexity).
      0 references

      Identifiers