[petsc-dev] Suggestions for MatProductCreate()

[Satish Balay]

On Mon, 23 Mar 2020, hzhang--- via petsc-dev wrote:

> Lisandro:
> 
> > * Please consider fixing MatProductCreate(A,B,C,&D) to take ownership
> > (that is, increase reference count) of the A,B, and the (optional) C
> > matrices provided as arguments. Otherwise it is way easy to get into the
> > dangling pointer trap.
> >
> Can you give me a simple example of " get into the dangling pointer trap"?
> We do not use reference count to keep track of A, B for Mat-Mat operations
> in the current and previous versions.
> 
> >
> > * A thing also missing in the new API is a way to "cleanup" the A,B,C
> > references, something MatProductReset(D) to get rid of (deallocates) the
> > internal "product" context, thus removing  from D the references to A,B,C.
> > This would be useful if you just want to compute JUST the symbolic product,
> > I'm using that in some code to compute the nonzero pattern of A^2.
> >
> Again, giving an example would help me understand. If you just want
> the  symbolic product, you can call
> MatProductCreate()
> MatProductSetType()
> MatProductSetFromOptions()
> MatProductSymbolic().
> This is equivalent to previous MatMatMultSymbolic(), and is used in some
> routines of PETSc.
> 
> >
> > * It should be also considered to provide backward compatibility
> > PETSC_DEPRECATED calls to the previous MatMatMultSymbolic()
> > and MatMatMultNumeric(). It looks like it would be trivial to do, though I
> > may be getting it wrong because I have not looked at all the details.
> >
>  MatMatMultSymbolic/Numeric() are not recommended for users, and few
> developers ever used them.  I only see one or two PETSc subroutines call
> them. I do not think we need provide backward compatibility
> PETSC_DEPRECATED calls for 6 pairs of such routines.

If there is a simple map from old API to new API [just new names, or reorder arguments] - we should include the old API with PETSC_DEPRECATED

Alternative is PETSC_DEPRECATED with some error statement? Jed will know better..


Satish

> 
> Hong
>