[petsc-dev] Counting Fortran function pointers

Barry Smith bsmith at mcs.anl.gov
Wed Dec 5 08:03:16 CST 2012


On Dec 4, 2012, at 11:32 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:

> You understand this just fine. ;-)
> 
> The dmdasnes, dmdats, etc depends on DM, not vice-versa. But if you have an enum for DM, then DM contains knowledge about users of DM, and you have to modify DM when something outside needs it.

   I think you mean that DMDASNESSetFunctionLocal() sets a function pointer specific to DMSNES_DA (I don't think you mean to talk about plain "DM" here (yes I use the wrong object for a fortran function pointer recently but that was because I was lazy).

    Anyways, the cure is that each PetscObject has two fortran pointer arrays; the current one that handles the methods for the base class of the PetscObject (for example DMSNESSetFunction()) and the other specific for the subclass (for example DMDASNESSetFunctionLocal()). When a subclass gets changed with XXXSetType() the second fortran pointer array is deleted so you don't get "left over" wrong function pointers. Each subclass manages its own enums for the subclass so the base class fortran stuff doesn't need to know about the subclasses.

   Does that work?

    Barry

> 
> Note that this also sucks if we have users implementing DM in their own codes.
> 
> On Dec 4, 2012 8:29 PM, "Barry Smith" <bsmith at mcs.anl.gov> wrote:
> 
> On Dec 4, 2012, at 10:22 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> 
> > That is what I implemented for TS. The problem is false dependency in the sense that the enum includes extra stuff. For example, DM function pointers are needed by dmdasnes.
> 
>    I don't understand this.
> 
> > With dynamic keys, we could avoid that spurious dependency.
> 
>    Not worth the extra initialization code.
> >
> > On Dec 4, 2012 8:14 PM, "Barry Smith" <bsmith at mcs.anl.gov> wrote:
> >
> > On Dec 4, 2012, at 4:43 PM, Jed Brown <jedbrown at mcs.anl.gov> wrote:
> >
> > > As usual, anything that is duplicated and not checked by the compiler is broken.
> > >
> > > $ grep PetscObjectAllocateFortranPointers src/**/*.c
> > > src/dm/impls/da/ftn-custom/zda2f.c:  PetscObjectAllocateFortranPointers(*da,6);
> > > src/dm/impls/da/ftn-custom/zda2f.c:  PetscObjectAllocateFortranPointers(*da,6);
> > > src/dm/impls/shell/ftn-custom/zdmshellf.c:  PetscObjectAllocateFortranPointers(*dm,2);
> > > src/dm/impls/shell/ftn-custom/zdmshellf.c:  PetscObjectAllocateFortranPointers(*dm,2);
> > >
> > > Note that changing the type does not reset the function pointers, thus having a DMSHELL, calling DMSetType(dm,DMDA), and then setting a DMDA local function will cause memory corruption.
> >
> > > I cannot express how much I hate this system. The full-blown solution is that for each type, we register a (global) token which is the index of that function pointer. That doesn't have any false dependencies, but is more "initialize" code.
> > >
> > > An alternative, used in the TS and KSP code below, is to have a common enum that lists all the Fortran functions. It's a false header dependency, but not a binary dependency.
> > > What should we do? The current state is a disaster.
> >
> >
> >    Each object gets an enum for each fortran implementable method in that object, names of the enums reflect the object and method name. The final enum for each object is the count of the number of the enums and is used in the AllocationFortranPointers.  Yes, some brain numbing work to set up initially but then pretty clean and extensible for the future (since the final enum is the count we don't have to worry about always fixing the count in the AllocateFunctionPointers() each time a new method is added).
> >
> >    Barry
> >
> > >
> > > src/ksp/ksp/impls/gmres/fgmres/ftn-custom/zmodpcff.c:  PetscObjectAllocateFortranPointers(*ksp,3);
> > > src/ksp/ksp/interface/ftn-custom/zitfuncf.c:  PetscObjectAllocateFortranPointers(*ksp,FTN_MAX);
> > > src/ksp/ksp/interface/ftn-custom/zitfuncf.c:  PetscObjectAllocateFortranPointers(*ksp,FTN_MAX);
> > > src/ksp/pc/impls/mg/ftn-custom/zmgfuncf.c:    PetscObjectAllocateFortranPointers(*mat,1);
> > > src/ksp/pc/impls/shell/ftn-custom/zshellpcf.c:  PetscObjectAllocateFortranPointers(*pc,5);
> > > src/ksp/pc/impls/shell/ftn-custom/zshellpcf.c:  PetscObjectAllocateFortranPointers(*pc,5);
> > > src/ksp/pc/impls/shell/ftn-custom/zshellpcf.c:  PetscObjectAllocateFortranPointers(*pc,5);
> > > src/ksp/pc/impls/shell/ftn-custom/zshellpcf.c:  PetscObjectAllocateFortranPointers(*pc,5);
> > > src/ksp/pc/impls/shell/ftn-custom/zshellpcf.c:  PetscObjectAllocateFortranPointers(*pc,5);
> > > src/mat/impls/mffd/ftn-custom/zmffdf.c:  PetscObjectAllocateFortranPointers(*mat,2);
> > > src/mat/impls/shell/ftn-custom/zshellf.c:  PetscObjectAllocateFortranPointers(*mat,11);
> > > src/mat/interface/ftn-custom/zmatrixf.c:  PetscObjectAllocateFortranPointers(*sp,1);
> > > src/snes/interface/ftn-custom/zsnesf.c:  PetscObjectAllocateFortranPointers(*snes,14);
> > > src/snes/interface/ftn-custom/zsnesf.c:  PetscObjectAllocateFortranPointers(*snes,14);
> > > src/snes/interface/ftn-custom/zsnesf.c:  PetscObjectAllocateFortranPointers(*snes,14);
> > > src/snes/interface/ftn-custom/zsnesf.c:  PetscObjectAllocateFortranPointers(*snes,14);
> > > src/snes/interface/ftn-custom/zsnesf.c:  PetscObjectAllocateFortranPointers(*snes,14);
> > > src/snes/linesearch/impls/shell/ftn-custom/zlinesearchshellf.c:  PetscObjectAllocateFortranPointers(*linesearch,3);
> > > src/snes/linesearch/interface/ftn-custom/zlinesearchf.c:  PetscObjectAllocateFortranPointers(*linesearch,3);
> > > src/snes/linesearch/interface/ftn-custom/zlinesearchf.c:  PetscObjectAllocateFortranPointers(*linesearch,3);
> > > src/sys/draw/utils/ftn-custom/zzoomf.c:  PetscObjectAllocateFortranPointers(*draw,1);
> > > src/ts/interface/ftn-custom/ztsf.c:  PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> > > src/ts/interface/ftn-custom/ztsf.c:  PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> > > src/ts/interface/ftn-custom/ztsf.c:    PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> > > src/ts/interface/ftn-custom/ztsf.c:    PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> > > src/ts/interface/ftn-custom/ztsf.c:  PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> > > src/ts/interface/ftn-custom/ztsf.c:  PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> > > src/ts/interface/ftn-custom/ztsf.c:  PetscObjectAllocateFortranPointers(*ts,OUR_COUNT);
> >
> 




More information about the petsc-dev mailing list