February 2013 - petsc-dev - mailman3.cels.anl.gov

Re: [petsc-dev] questions on new include organization
by Barry Smith 12 Feb '13

12 Feb '13

On Feb 12, 2013, at 11:27 PM, Jed Brown <jedbrown(a)mcs.anl.gov> wrote: > > On Tue, Feb 12, 2013 at 11:24 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > Then a PetscSection is a collection of lists of integers. An IS is a single list of integers. So let's make IS be a collection of lists of integers (what was previously a PetscSection) and the current use of IS becomes a special case. > > Then what does ISGetIndices() become? Then would depend on how we would define how the grouping is handled and how the GetIndices would be used. Some possibilities 1) ISGetIndices(is,int group_number, &indices) gives the indices for a group or 2) ISGetIndices(is,&indices) simply returns all the indices for all the groups

1 0

Re: [petsc-dev] questions on new include organization
by Jed Brown 12 Feb '13

12 Feb '13

On Tue, Feb 12, 2013 at 11:24 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > Then a PetscSection is a collection of lists of integers. An IS is a > single list of integers. So let's make IS be a collection of lists of > integers (what was previously a PetscSection) and the current use of IS > becomes a special case. Then what does ISGetIndices() become?

1 0

Re: [petsc-dev] questions on new include organization
by Jed Brown 12 Feb '13

12 Feb '13

On Tue, Feb 12, 2013 at 11:17 PM, Matthew Knepley <knepley(a)gmail.com> wrote: > Yes. PetscSection represents groups of numbers, and you can see the > grouping. > IS is a map from (int) -> (int), but PetscSection is a map from (int) -> (int,int) where the pair is interpreted to mean an interval of non-negative extent. If it's any use, an IS could be thought of as the special case (int) -> (int,1). > > >> What queries "do not just return members of this list" in PETSc >> section? (That is queries we wish to retain anyways). > > > For example, you can ask "how many items are in group p?". This is > crucial, and the whole point of the structure. > I don't see how you can recover that information from just the list of dof > numbers. >

1 0

Re: [petsc-dev] questions on new include organization
by Barry Smith 12 Feb '13

12 Feb '13

On Feb 12, 2013, at 11:17 PM, Matthew Knepley <knepley(a)gmail.com> wrote: > Regardless of these two, PetscSection does not belong in the vector subdirectory and most definitely not in the impls/seq directory, so even if in the end it does not become exactly an IS it more belongs in the is directory then in the vec directory > > That is convincing. > > > c) I believe the claim will be, "I can always think of the output as a list of integers", however that is a very broad category, > > which also includes graphs, meshes, etc. Internally this looks nothing like a list of numbers, and the current queries > > do not just return members of this list. > > a graph and mesh most definitely cannot be thought of as a list of integers since they contain information about relationships BETWEEN items. Does PetscSection contain such additional information? > > Yes. PetscSection represents groups of numbers, and you can see the grouping. > > What queries "do not just return members of this list" in PETSc section? (That is queries we wish to retain anyways). > > For example, you can ask "how many items are in group p?". This is crucial, and the whole point of the structure. > I don't see how you can recover that information from just the list of dof numbers. Then a PetscSection is a collection of lists of integers. An IS is a single list of integers. So let's make IS be a collection of lists of integers (what was previously a PetscSection) and the current use of IS becomes a special case. Barry > > Matt > > > Barry > > > > > Matt > > > > > > 2) matimpl.h really needs vecimpl.h? > > > > #define __MATIMPL_H > > > > #include <petscmat.h> > > #include <petsc-private/petscimpl.h> > > #include <petsc-private/vecimpl.h> > > > > Doesn't seem to so I've removed the vecimpl.h and pushed > > > > Seems there may be a few of these unneeded dependencies hanging around, I'll try to hunt them down. > > > > Barry > > > > > > > > > > -- > > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. > > -- Norbert Wiener > > > > > -- > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. > -- Norbert Wiener

1 0

Re: [petsc-dev] questions on new include organization
by Barry Smith 12 Feb '13

12 Feb '13

On Feb 12, 2013, at 11:13 PM, Satish Balay <balay(a)mcs.anl.gov> wrote: > On Tue, 12 Feb 2013, Barry Smith wrote: > >> 2) matimpl.h really needs vecimpl.h? >> >> #define __MATIMPL_H >> >> #include <petscmat.h> >> #include <petsc-private/petscimpl.h> >> #include <petsc-private/vecimpl.h> >> >> Doesn't seem to so I've removed the vecimpl.h and pushed >> >> Seems there may be a few of these unneeded dependencies hanging around, I'll try to hunt them down. > > https://bitbucket.org/petsc/petsc-dev/commits/2e1692285b5bc97753e7e4964d9a3… > > To remove vecimpl.h [from matimpl.h] - we have to move PetscCUSPFlag > to a more appropriate location. Thanks. I've moved it to petscimpl.h > > Satish

1 0

Re: [petsc-dev] questions on new include organization
by Matthew Knepley 12 Feb '13

12 Feb '13

On Wed, Feb 13, 2013 at 12:08 AM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > On Feb 12, 2013, at 10:56 PM, Matthew Knepley <knepley(a)gmail.com> wrote: > > > On Tue, Feb 12, 2013 at 11:41 PM, Barry Smith <bsmith(a)mcs.anl.gov> > wrote: > > > > 1) Can we move the section data structures from vecimpl.h to isimpl.h? > in preparation for its refactoring and why is PetscSectionCreate in > section.c defined in src/vec/vec/impls/seq? Can vsection.c be split in > two with the only the "Vec" related routines kept in the Vec directory > > > > All the above makes sense if you agree with this designation as an IS. > Can we go over this again to make sure it is correct? > > > > a) PetscSection does a job very similar to PetscLayout, but Layout is > not an IS > > > > b) The internals of PetscSection look like two ISes, and a list of > embedded PetscSections. This does not sound like an IS > > > Regardless of these two, PetscSection does not belong in the vector > subdirectory and most definitely not in the impls/seq directory, so even > if in the end it does not become exactly an IS it more belongs in the is > directory then in the vec directory That is convincing. > > c) I believe the claim will be, "I can always think of the output as a > list of integers", however that is a very broad category, > > which also includes graphs, meshes, etc. Internally this looks > nothing like a list of numbers, and the current queries > > do not just return members of this list. > > a graph and mesh most definitely cannot be thought of as a list of > integers since they contain information about relationships BETWEEN items. > Does PetscSection contain such additional information? > Yes. PetscSection represents groups of numbers, and you can see the grouping. > What queries "do not just return members of this list" in PETSc > section? (That is queries we wish to retain anyways). For example, you can ask "how many items are in group p?". This is crucial, and the whole point of the structure. I don't see how you can recover that information from just the list of dof numbers. Matt > > Barry > > > > > Matt > > > > > > 2) matimpl.h really needs vecimpl.h? > > > > #define __MATIMPL_H > > > > #include <petscmat.h> > > #include <petsc-private/petscimpl.h> > > #include <petsc-private/vecimpl.h> > > > > Doesn't seem to so I've removed the vecimpl.h and pushed > > > > Seems there may be a few of these unneeded dependencies hanging around, > I'll try to hunt them down. > > > > Barry > > > > > > > > > > -- > > What most experimenters take for granted before they begin their > experiments is infinitely more interesting than any results to which their > experiments lead. > > -- Norbert Wiener > > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener

1 0

Re: [petsc-dev] questions on new include organization
by Satish Balay 12 Feb '13

12 Feb '13

On Tue, 12 Feb 2013, Barry Smith wrote: > 2) matimpl.h really needs vecimpl.h? > > #define __MATIMPL_H > > #include <petscmat.h> > #include <petsc-private/petscimpl.h> > #include <petsc-private/vecimpl.h> > > Doesn't seem to so I've removed the vecimpl.h and pushed > > Seems there may be a few of these unneeded dependencies hanging around, I'll try to hunt them down. https://bitbucket.org/petsc/petsc-dev/commits/2e1692285b5bc97753e7e4964d9a3… To remove vecimpl.h [from matimpl.h] - we have to move PetscCUSPFlag to a more appropriate location. Satish

1 0

Re: [petsc-dev] questions on new include organization
by Barry Smith 12 Feb '13

12 Feb '13

On Feb 12, 2013, at 10:56 PM, Matthew Knepley <knepley(a)gmail.com> wrote: > On Tue, Feb 12, 2013 at 11:41 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > 1) Can we move the section data structures from vecimpl.h to isimpl.h? in preparation for its refactoring and why is PetscSectionCreate in section.c defined in src/vec/vec/impls/seq? Can vsection.c be split in two with the only the "Vec" related routines kept in the Vec directory > > All the above makes sense if you agree with this designation as an IS. Can we go over this again to make sure it is correct? > > a) PetscSection does a job very similar to PetscLayout, but Layout is not an IS > > b) The internals of PetscSection look like two ISes, and a list of embedded PetscSections. This does not sound like an IS > Regardless of these two, PetscSection does not belong in the vector subdirectory and most definitely not in the impls/seq directory, so even if in the end it does not become exactly an IS it more belongs in the is directory then in the vec directory > c) I believe the claim will be, "I can always think of the output as a list of integers", however that is a very broad category, > which also includes graphs, meshes, etc. Internally this looks nothing like a list of numbers, and the current queries > do not just return members of this list. a graph and mesh most definitely cannot be thought of as a list of integers since they contain information about relationships BETWEEN items. Does PetscSection contain such additional information? What queries "do not just return members of this list" in PETSc section? (That is queries we wish to retain anyways). Barry > > Matt > > > 2) matimpl.h really needs vecimpl.h? > > #define __MATIMPL_H > > #include <petscmat.h> > #include <petsc-private/petscimpl.h> > #include <petsc-private/vecimpl.h> > > Doesn't seem to so I've removed the vecimpl.h and pushed > > Seems there may be a few of these unneeded dependencies hanging around, I'll try to hunt them down. > > Barry > > > > > -- > What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. > -- Norbert Wiener

1 0

Re: [petsc-dev] questions on new include organization
by Matthew Knepley 12 Feb '13

12 Feb '13

On Tue, Feb 12, 2013 at 11:41 PM, Barry Smith <bsmith(a)mcs.anl.gov> wrote: > > 1) Can we move the section data structures from vecimpl.h to isimpl.h? in > preparation for its refactoring and why is PetscSectionCreate in section.c > defined in src/vec/vec/impls/seq? Can vsection.c be split in two with the > only the "Vec" related routines kept in the Vec directory All the above makes sense if you agree with this designation as an IS. Can we go over this again to make sure it is correct? a) PetscSection does a job very similar to PetscLayout, but Layout is not an IS b) The internals of PetscSection look like two ISes, and a list of embedded PetscSections. This does not sound like an IS c) I believe the claim will be, "I can always think of the output as a list of integers", however that is a very broad category, which also includes graphs, meshes, etc. Internally this looks nothing like a list of numbers, and the current queries do not just return members of this list. Matt > > 2) matimpl.h really needs vecimpl.h? > > #define __MATIMPL_H > > #include <petscmat.h> > #include <petsc-private/petscimpl.h> > #include <petsc-private/vecimpl.h> > > Doesn't seem to so I've removed the vecimpl.h and pushed > > Seems there may be a few of these unneeded dependencies hanging around, > I'll try to hunt them down. > > Barry > > -- What most experimenters take for granted before they begin their experiments is infinitely more interesting than any results to which their experiments lead. -- Norbert Wiener

1 0

questions on new include organization
by Barry Smith 12 Feb '13

12 Feb '13

1) Can we move the section data structures from vecimpl.h to isimpl.h? in preparation for its refactoring and why is PetscSectionCreate in section.c defined in src/vec/vec/impls/seq? Can vsection.c be split in two with the only the "Vec" related routines kept in the Vec directory? 2) matimpl.h really needs vecimpl.h? #define __MATIMPL_H #include <petscmat.h> #include <petsc-private/petscimpl.h> #include <petsc-private/vecimpl.h> Doesn't seem to so I've removed the vecimpl.h and pushed Seems there may be a few of these unneeded dependencies hanging around, I'll try to hunt them down. Barry

1 0