June 2012 - petsc-dev - mailman3.cels.anl.gov

Re: [petsc-dev] PetscLogView vs. -log_summary
by Matthew Knepley 21 Jun '12

21 Jun '12

On Thu, Jun 21, 2012 at 2:08 PM, Blaise Bourdin <bourdin(a)lsu.edu> wrote: > > On Jun 21, 2012, at 2:55 PM, Matthew Knepley wrote: > > On Thu, Jun 21, 2012 at 1:11 PM, Blaise Bourdin <bourdin(a)math.lsu.edu>wrote: > >> Hi, >> >> PetscLogView does not display the same amount of information as running >> with -log_summary. In the "summary of stages" section, the former contains >> only information on the main stage while the later contains much more. >> >> Add the following block at the end of src/snes/examples/tutorials/ex1.c, >> for instance, to see what I mean >> PetscViewer viewer; >> ierr = >> PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.py",&viewer);CHKERRQ(ierr); >> ierr = PetscLogViewPython(viewer);CHKERRQ(ierr); >> ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); >> ierr = >> PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.txt",&viewer);CHKERRQ(ierr); >> ierr = PetscLogView(viewer);CHKERRQ(ierr); >> ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); >> ierr = >> PetscLogPrintDetailed(PETSC_COMM_WORLD,"ex1-logdetailed.txt");CHKERRQ(ierr); >> >> Is this a bug or a feature? Is there a way to get the same output as >> -log_summary in a file, without printing it to stdout? > > > -log_summary just calls PetscLogView(). I am not sure how there could be a > problem. I will run it. However, can you > just use > > -log_summary <filename> > > > That works, yes. Is the issue that stages / events need to be activated a > priori, which is not possible when PetscLogView is only called at the end > of the program? > Crap. If you do not provide -log_summary, you need PetscLogBegin() at the beginning. Matt > The caveat is that I am assuming that when using log_summary, the log is > written as part of PetscFinalize, so that if a job is killed by the > sceduler or the user, there is no way to get an idea of the performances. I > like to print the log every so often. > > Blaise > > -- > Department of Mathematics and Center for Computation & Technology > Louisiana State University, Baton Rouge, LA 70803, USA > Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 > http://www.math.lsu.edu/~bourdin > > > > > > > > -- 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] PetscLogView vs. -log_summary
by Blaise Bourdin 21 Jun '12

21 Jun '12

On Jun 21, 2012, at 2:55 PM, Matthew Knepley wrote: > On Thu, Jun 21, 2012 at 1:11 PM, Blaise Bourdin <bourdin(a)math.lsu.edu> wrote: > Hi, > > PetscLogView does not display the same amount of information as running with -log_summary. In the "summary of stages" section, the former contains only information on the main stage while the later contains much more. > > Add the following block at the end of src/snes/examples/tutorials/ex1.c, for instance, to see what I mean > PetscViewer viewer; > ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.py",&viewer);CHKERRQ(ierr); > ierr = PetscLogViewPython(viewer);CHKERRQ(ierr); > ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); > ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.txt",&viewer);CHKERRQ(ierr); > ierr = PetscLogView(viewer);CHKERRQ(ierr); > ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); > ierr = PetscLogPrintDetailed(PETSC_COMM_WORLD,"ex1-logdetailed.txt");CHKERRQ(ierr); > > Is this a bug or a feature? Is there a way to get the same output as -log_summary in a file, without printing it to stdout? > > -log_summary just calls PetscLogView(). I am not sure how there could be a problem. I will run it. However, can you > just use > > -log_summary <filename> That works, yes. Is the issue that stages / events need to be activated a priori, which is not possible when PetscLogView is only called at the end of the program? The caveat is that I am assuming that when using log_summary, the log is written as part of PetscFinalize, so that if a job is killed by the sceduler or the user, there is no way to get an idea of the performances. I like to print the log every so often. Blaise -- Department of Mathematics and Center for Computation & Technology Louisiana State University, Baton Rouge, LA 70803, USA Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 http://www.math.lsu.edu/~bourdin

1 0

Re: [petsc-dev] PetscLogView vs. -log_summary
by Matthew Knepley 21 Jun '12

21 Jun '12

On Thu, Jun 21, 2012 at 1:11 PM, Blaise Bourdin <bourdin(a)math.lsu.edu>wrote: > Hi, > > PetscLogView does not display the same amount of information as running > with -log_summary. In the "summary of stages" section, the former contains > only information on the main stage while the later contains much more. > > Add the following block at the end of src/snes/examples/tutorials/ex1.c, > for instance, to see what I mean > PetscViewer viewer; > ierr = > PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.py",&viewer);CHKERRQ(ierr); > ierr = PetscLogViewPython(viewer);CHKERRQ(ierr); > ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); > ierr = > PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.txt",&viewer);CHKERRQ(ierr); > ierr = PetscLogView(viewer);CHKERRQ(ierr); > ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); > ierr = > PetscLogPrintDetailed(PETSC_COMM_WORLD,"ex1-logdetailed.txt");CHKERRQ(ierr); > > Is this a bug or a feature? Is there a way to get the same output as > -log_summary in a file, without printing it to stdout? -log_summary just calls PetscLogView(). I am not sure how there could be a problem. I will run it. However, can you just use -log_summary <filename> Matt > > Blaise > -- > Department of Mathematics and Center for Computation & Technology > Louisiana State University, Baton Rouge, LA 70803, USA > Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 > http://www.math.lsu.edu/~bourdin > > > > > > > > -- 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] Modifying code in seqcusp folder
by Matthew Knepley 21 Jun '12

21 Jun '12

On Thu, Jun 21, 2012 at 1:48 PM, Chekuri Choudary <cchoudary(a)rnet-tech.com>wrote: > Typically, when I make changes to petsc code, I do a make lib in the > folder and then recompile my code. That doesn’t seem to be working in > seqcusp folder. I do a make lib and nothing happens. Can somebody please > help?? > Yes, the check for stale CUDA code was broken. Satish, did this get fixed or is it still broken? Matt -- 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

Modifying code in seqcusp folder
by Chekuri Choudary 21 Jun '12

21 Jun '12

Typically, when I make changes to petsc code, I do a make lib in the folder and then recompile my code. That doesn't seem to be working in seqcusp folder. I do a make lib and nothing happens. Can somebody please help??

1 0

PetscLogView vs. -log_summary
by Blaise Bourdin 21 Jun '12

21 Jun '12

Hi, PetscLogView does not display the same amount of information as running with -log_summary. In the "summary of stages" section, the former contains only information on the main stage while the later contains much more. Add the following block at the end of src/snes/examples/tutorials/ex1.c, for instance, to see what I mean PetscViewer viewer; ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.py",&viewer);CHKERRQ(ierr); ierr = PetscLogViewPython(viewer);CHKERRQ(ierr); ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); ierr = PetscViewerASCIIOpen(PETSC_COMM_WORLD,"ex1-log.txt",&viewer);CHKERRQ(ierr); ierr = PetscLogView(viewer);CHKERRQ(ierr); ierr = PetscViewerDestroy(&viewer);CHKERRQ(ierr); ierr = PetscLogPrintDetailed(PETSC_COMM_WORLD,"ex1-logdetailed.txt");CHKERRQ(ierr); Is this a bug or a feature? Is there a way to get the same output as -log_summary in a file, without printing it to stdout? Blaise -- Department of Mathematics and Center for Computation & Technology Louisiana State University, Baton Rouge, LA 70803, USA Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 http://www.math.lsu.edu/~bourdin

1 0

Re: [petsc-dev] Getting "from" and "to" IS from a VecScatter
by Blaise Bourdin 21 Jun '12

21 Jun '12

On Jun 21, 2012, at 7:42 AM, Matthew Knepley wrote: > On Wed, Jun 20, 2012 at 8:40 PM, Blaise Bourdin <bourdin(a)lsu.edu> wrote: > Hi, > > I want to change the way I handle boundary conditions in my finite elements code: > Right now, I skip the rows and columns during the matrices assembly, using a flag (a SectionInt). Using an IS denoting all dof with a BC and MatZeroRowsColumns, MatZeroRowsColumnsIS or MatZeroRowsColumnsLocal would make much more sense. > > The difficulty comes up when applying a Dirichlet boundary condition on only one of the dof associated with a vertex (roller boundary conditions in elasticity, for instance). What I can easily get from Sieve is the point number associated with a vertex, for instance, from which I can obtain the _local_ index of the dof associated with a given component of a field at this point. For vectors, I can then set values on the local Vec or the Section. For matrices, though I can't just use MatGetLocalSubMatrix since I need also to modify the out of diagonal block terms. > > I guess I need the LocalToGlobalMapping or to renumber the IS describing my BC and convert it to global indices. I am sure that this information is somewhere in Sieve (in the overlap, I guess), but I was thinking that since getting a Local to Global scatter is trivial, there may be a way to obtain the mapping or renumbering the IS from the scatter. > > Now that I think about it more, the right way is most likely to get the information I need from Sieve. > > Yes, in the new version there is jsut a LocalToGlobalMapping. In C++, you can use the GlobalNumbering, > > numbering->getIndex(point) + dof offset And for the record, in the old version, where a DMMesh can be associated with several data layout, the following seems to work (adapted from DMCreateLocalToGlobalMapping_Mesh): Obj<PETSC_MESH_TYPE::real_section_type> s; ALE::Obj<PETSC_MESH_TYPE> m; ierr = DMMeshGetMesh(dm, m);CHKERRQ(ierr); ierr = SectionRealGetSection(vsec,s);CHKERRQ(ierr); const ALE::Obj<PETSC_MESH_TYPE::order_type>& globalOrder = m->getFactory()->getGlobalOrder(m, s->getName(), s); offset = globalOrder->getIndex(point) + dof ; Blaise -- Department of Mathematics and Center for Computation & Technology Louisiana State University, Baton Rouge, LA 70803, USA Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 http://www.math.lsu.edu/~bourdin

1 0

Re: [petsc-dev] Getting "from" and "to" IS from a VecScatter
by Matthew Knepley 21 Jun '12

21 Jun '12

On Wed, Jun 20, 2012 at 8:40 PM, Blaise Bourdin <bourdin(a)lsu.edu> wrote: > Hi, > > I want to change the way I handle boundary conditions in my finite > elements code: > Right now, I skip the rows and columns during the matrices assembly, using > a flag (a SectionInt). Using an IS denoting all dof with a BC and > MatZeroRowsColumns, MatZeroRowsColumnsIS or MatZeroRowsColumnsLocal would > make much more sense. > > The difficulty comes up when applying a Dirichlet boundary condition on > only one of the dof associated with a vertex (roller boundary conditions in > elasticity, for instance). What I can easily get from Sieve is the point > number associated with a vertex, for instance, from which I can obtain the > _local_ index of the dof associated with a given component of a field at > this point. For vectors, I can then set values on the local Vec or the > Section. For matrices, though I can't just use MatGetLocalSubMatrix since I > need also to modify the out of diagonal block terms. > > I guess I need the LocalToGlobalMapping or to renumber the IS describing > my BC and convert it to global indices. I am sure that this information is > somewhere in Sieve (in the overlap, I guess), but I was thinking that since > getting a Local to Global scatter is trivial, there may be a way to obtain > the mapping or renumbering the IS from the scatter. > > Now that I think about it more, the right way is most likely to get the > information I need from Sieve. Yes, in the new version there is jsut a LocalToGlobalMapping. In C++, you can use the GlobalNumbering, numbering->getIndex(point) + dof offset Matt > > Blaise > > > On Jun 20, 2012, at 7:23 PM, Barry Smith wrote: > > > > > Why do you need them? Maybe there is a better way? > > > > > > On Jun 20, 2012, at 6:19 PM, Blaise Bourdin wrote: > > > >> Hi, > >> > >> I need to reconstruct the IS that were used to build a VecScatter > between a local and a global Vec. > >> > >> From what I see in VecScatterView_MPI (vpscat.c:13), it looks like I > can get the indices from from->local.vslots[i],to->local.vslots[i], i=0 .. > i<to->local.n where from and to are obtained with > >> VecScatter_MPI_General *to=(VecScatter_MPI_General*)ctx->todata; > >> VecScatter_MPI_General *from=(VecScatter_MPI_General*)ctx->fromdata; > >> > >> From there, constructing the IS is trivial. Is there a way that does > not involve poking inside the VecScatter data structure? > >> > >> Blaise > >> > >> > >> -- > >> Department of Mathematics and Center for Computation & Technology > >> Louisiana State University, Baton Rouge, LA 70803, USA > >> Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 > http://www.math.lsu.edu/~bourdin > >> > >> > >> > >> > >> > >> > >> > > > > -- > Department of Mathematics and Center for Computation & Technology > Louisiana State University, Baton Rouge, LA 70803, USA > Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 > http://www.math.lsu.edu/~bourdin > > > > > > > > -- 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] Getting "from" and "to" IS from a VecScatter
by Blaise Bourdin 20 Jun '12

20 Jun '12

Hi, I want to change the way I handle boundary conditions in my finite elements code: Right now, I skip the rows and columns during the matrices assembly, using a flag (a SectionInt). Using an IS denoting all dof with a BC and MatZeroRowsColumns, MatZeroRowsColumnsIS or MatZeroRowsColumnsLocal would make much more sense. The difficulty comes up when applying a Dirichlet boundary condition on only one of the dof associated with a vertex (roller boundary conditions in elasticity, for instance). What I can easily get from Sieve is the point number associated with a vertex, for instance, from which I can obtain the _local_ index of the dof associated with a given component of a field at this point. For vectors, I can then set values on the local Vec or the Section. For matrices, though I can't just use MatGetLocalSubMatrix since I need also to modify the out of diagonal block terms. I guess I need the LocalToGlobalMapping or to renumber the IS describing my BC and convert it to global indices. I am sure that this information is somewhere in Sieve (in the overlap, I guess), but I was thinking that since getting a Local to Global scatter is trivial, there may be a way to obtain the mapping or renumbering the IS from the scatter. Now that I think about it more, the right way is most likely to get the information I need from Sieve. Blaise On Jun 20, 2012, at 7:23 PM, Barry Smith wrote: > > Why do you need them? Maybe there is a better way? > > > On Jun 20, 2012, at 6:19 PM, Blaise Bourdin wrote: > >> Hi, >> >> I need to reconstruct the IS that were used to build a VecScatter between a local and a global Vec. >> >> From what I see in VecScatterView_MPI (vpscat.c:13), it looks like I can get the indices from from->local.vslots[i],to->local.vslots[i], i=0 .. i<to->local.n where from and to are obtained with >> VecScatter_MPI_General *to=(VecScatter_MPI_General*)ctx->todata; >> VecScatter_MPI_General *from=(VecScatter_MPI_General*)ctx->fromdata; >> >> From there, constructing the IS is trivial. Is there a way that does not involve poking inside the VecScatter data structure? >> >> Blaise >> >> >> -- >> Department of Mathematics and Center for Computation & Technology >> Louisiana State University, Baton Rouge, LA 70803, USA >> Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 http://www.math.lsu.edu/~bourdin >> >> >> >> >> >> >> > -- Department of Mathematics and Center for Computation & Technology Louisiana State University, Baton Rouge, LA 70803, USA Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 http://www.math.lsu.edu/~bourdin

1 0

Re: [petsc-dev] Getting "from" and "to" IS from a VecScatter
by Barry Smith 20 Jun '12

20 Jun '12

Why do you need them? Maybe there is a better way? On Jun 20, 2012, at 6:19 PM, Blaise Bourdin wrote: > Hi, > > I need to reconstruct the IS that were used to build a VecScatter between a local and a global Vec. > > From what I see in VecScatterView_MPI (vpscat.c:13), it looks like I can get the indices from from->local.vslots[i],to->local.vslots[i], i=0 .. i<to->local.n where from and to are obtained with > VecScatter_MPI_General *to=(VecScatter_MPI_General*)ctx->todata; > VecScatter_MPI_General *from=(VecScatter_MPI_General*)ctx->fromdata; > > From there, constructing the IS is trivial. Is there a way that does not involve poking inside the VecScatter data structure? > > Blaise > > > -- > Department of Mathematics and Center for Computation & Technology > Louisiana State University, Baton Rouge, LA 70803, USA > Tel. +1 (225) 578 1612, Fax +1 (225) 578 4276 http://www.math.lsu.edu/~bourdin > > > > > > >

1 0