recent changes to ILUDT and matrix factorization API

[Barry Smith]

    Lisandro,

      The LUDT is currently work in progress and will change as you  
suggest once it is close
to being done.

       Regarding in place factorization. For dense matrices in place  
makes perfect sense and
in fact is commonly (perhaps overly) used in the NA community.

    Barry

On Apr 28, 2009, at 12:07 PM, Lisandro Dalcin wrote:

> In a recent commit, MatILUDTFactorSymbolic() and
> MatILUDTFactorNumeric() were merged in MatILUDTFactor(). However,
> MatILUDTFactor() does not do itself the MatGetFactor() call (instead,
> that call is hardwired in PCSetUp_ILU()).
>
> IMHO, MatILUDTFactor() (or the type-specific implementations) should
> make the call to MatGetFactor(). This is like things work for
> MatLUFactor(), for example.
>
> Moreover, I believe PC ILU when "ilu->usedt" is now broken for the
> cases "pc->flag != SAME_NONZERO_PATTERN" and "!ilu->reusefill",
> because of the MatDestroy() calls.
>
> Finally I would like to point that I'm somewhat -1 in the whole change
> merging the calls. Let's take for example LU factorization.
>
> A) If you want to perform LU factorization with "in-place" semantics,
> you call MatLUFactor(A). In this case, the "in-place" is just
> semantics, as the operation is always truly inplace (certainly not for
> SeqAIJ, for example).
>
> B) If you want to perform LU factorization with "out-of-place"
> semantics, you call MatGetFactor(A,&fact) +
> MatLUFactorSymbolic(fact,A) + MatLUFactorNumeric(fact,A).
>
> I do not see why ILUDT (or even ILU(n) with n>1) should be treated
> differently. The only one that could be special cased is ILU(0),
> because it can be done truly in-place if the user request it.
>
> Perhaps other option is remove all calls like (A), just because truly
> "in-place" cannot be done in practice, and special case ILU(0) and
> ICC(0). I mean, the only calls with "in-place" semantics should be
> MatILUFactor() and MatICCFactor(), but the fail if the levels of fill
> is not zero.
>
>
>
>
> -- 
> Lisandro Dalcín
> ---------------
> Centro Internacional de Métodos Computacionales en Ingeniería (CIMEC)
> Instituto de Desarrollo Tecnológico para la Industria Química (INTEC)
> Consejo Nacional de Investigaciones Científicas y Técnicas (CONICET)
> PTLC - Güemes 3450, (3000) Santa Fe, Argentina
> Tel/Fax: +54-(0)342-451.1594