Wrap-around partitioning for block bidiagonal linear systems (Q3840701)

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 1187623
Language Label Description Also known as
default for all languages
No label defined
    English
    Wrap-around partitioning for block bidiagonal linear systems
    scientific article; zbMATH DE number 1187623

      Statements

      Wrap-around partitioning for block bidiagonal linear systems (English)
      0 references
      0 references
      0 references
      18 February 1999
      0 references
      vector algorithm
      0 references
      block bidiagonal linear systems
      0 references
      wrap-around partitioning
      0 references
      QR factorization
      0 references
      A stable vector algorithm for the solution of the block bidiagonal linear systems is obtained combining a wrap-around partitioning (reordering of the unknowns in a system of linear equations) and the QR factorization. A step of the wrap-around partitioning transforms a linear system into a system of the \(q\) (super)blocks of \(p\) (submatrix) elements each. Then the orthogonal transformations is considered for elimination of the \(q-1\) subdiagonal \(p\times p\) block matrices. Each subdiagonal block row in the current pivotal block column can be eliminated independently of the other block rows. This observation provides the basis for vectorizing the computation. The final block is again block bidiagonal in the structure and so it is possible to use the recursive implementation of the wrap-around partitioning. Organization of the wrap-around partitioning and elimination of the subdiagonal blocks are discussed. A determination of the optimum transformation parameters \(p\) and \(q\) is considered. Certain implementation details are discussed and some conclusions are drawn. Using a simple model for the vectorization overhead it is shown that small block sizes give the best performance.
      0 references

      Identifiers