Re: [eigen] new API for Cholmod

[ Thread Index | Date Index | More Archives ]

On 11/ 4/10 01:10 PM, Bill Greene wrote:
1. The ordering algorithm in a particular sparse matrix package
is likely to be useful with other sparse matrix factoring routines (i.e.
    the CSparse AMD would be useful with MUMPS or SPOOLES, etc.)
The Metis ordering package is often very effective and is distributed or
    recommended in several sparse matrix packages-- probably all
    sparse backends would find this useful.

2. Users who are particularly interested in performance, often like to
    experiment with different ordering algorithms to find the optimum
    one for their problem.

So its probably better to allow the user to separate the ordering step
from the analysis and factorization steps.

Excellent points.
>This still require some work to make it compatible with any Eigen's
>sparse representation (SparseMatrix, DynamicSparseMatrix, row-column
>major, with only the lower, or upper or both triangular part etc.).

Yes, I can see how this would be quite a bit of work. Did you consider
just stopping with an error message if the matrix is not a SparseMatrix
in the correct form-- the upper triangle for CSparse? As a user I don't
think that would be so objectionable. I could then choose to convert the
matrix to the correct format, myself, possibly avoiding the memory cost
of having two copies of the matrix.
Definitely.  CHOLMOD is a brilliant piece of software.  But, one
thing that has caught me with my pants down is that it will
silently translate your problem into AA' if you forget to tell it that A
is symmetric (and why is that anyways, then isn't the answer just A?).
Let's not have any of that.

Rodney Sparapani       Center for Patient Care and Outcomes Research
Sr. Biostatistician     
4 wheels good, 2 wheels better!   Medical College of Wisconsin (MCW)
WWLD?:  What Would Lombardi Do?   Milwaukee, WI, USA

Mail converted by MHonArc 2.6.19+