next up previous index
Next: LU Decomposition Up: Gauss-Jordan Elimination Previous: The Discussion

Gauss-Jordan and HPF

The algorithm sweeps down the matrix from the top left corner to the bottom right corner, leaving zero subdiagonal elements behind it.


 
Figure 5.1: Communication and computation in the various phases of the HPF Gaussian (from Foster)
\includegraphics*[width=5cm,angle=270]{img1000.ps}

What is parallel in the algorithm?

1.
MAXLOC: reduction operation on the row and column defined by the mask lpiv, then broadcast within that row and column
2.
scale factors require N-n independent operations within column icol
3.
scale factor and a pivot row value must be broadcast within each column and row respectively
4.
the reductions require ${\cal O}((N-n)^2)$ independent operations

Attributes of the computation:


next up previous index
Next: LU Decomposition Up: Gauss-Jordan Elimination Previous: The Discussion
Zdzislaw Meglicki
2001-02-26