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

Getting "from" and "to" IS from a VecScatter
by Blaise Bourdin 20 Jun '12

20 Jun '12

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

Re: [petsc-dev] urgent issues with DMDA_BOUNDARY_GHOSTED DMDA_BOUNDARY_MIRROR
by Barry Smith 20 Jun '12

20 Jun '12

My inclination is to remove the mirror enum and simply provide a utility function that updates the ghost points by mirroring in the local vector form that the user can call if they want mirroring? Barry On Jun 20, 2012, at 5:41 PM, Barry Smith wrote: > > What is the intended difference between DMDA_BOUNDARY_GHOSTED and DMDA_BOUNDARY_MIRROR? > > My understanding is that both cause ghost points around the physical boundary that can be indexed in the usual way (for example on the left side just indexing with -1 to get the first of the ghost points? Looking at the code but not testing it I assume that the correctly sized local vectors are created and the appropriate scatters are created that put things in the correct place despite these extra locations with calls to DMGlobalToLocal()? > > My guess is that whoever decided to have both of these was thinking that with DMDA_BOUNDARY_GHOSTED the user would fill up the ghost points themselves however they liked with whatever values they want whenever they want? Meanwhile with mirror PETSc would fill up the ghost points by copying the mirrored values from inside the real physical domain? Thus ghosted is more general. > > But when was it intended that the mirrored values be copied over? During the DMGlobalToLocal call or in a new call? (Note that this copying requires no parallel communication) > > Is there an example of using DMDA_BOUNDARY_GHOSTED where the ghosted values are set via mirroring? Should there be a utility to do that? > > We have someone who needs this mirroring (due to Jed's advice) but appear not to have any code to make that trivial. How shall we fix this? > > Thanks > > Barry >

1 0

urgent issues with DMDA_BOUNDARY_GHOSTED DMDA_BOUNDARY_MIRROR
by Barry Smith 20 Jun '12

20 Jun '12

What is the intended difference between DMDA_BOUNDARY_GHOSTED and DMDA_BOUNDARY_MIRROR? My understanding is that both cause ghost points around the physical boundary that can be indexed in the usual way (for example on the left side just indexing with -1 to get the first of the ghost points? Looking at the code but not testing it I assume that the correctly sized local vectors are created and the appropriate scatters are created that put things in the correct place despite these extra locations with calls to DMGlobalToLocal()? My guess is that whoever decided to have both of these was thinking that with DMDA_BOUNDARY_GHOSTED the user would fill up the ghost points themselves however they liked with whatever values they want whenever they want? Meanwhile with mirror PETSc would fill up the ghost points by copying the mirrored values from inside the real physical domain? Thus ghosted is more general. But when was it intended that the mirrored values be copied over? During the DMGlobalToLocal call or in a new call? (Note that this copying requires no parallel communication) Is there an example of using DMDA_BOUNDARY_GHOSTED where the ghosted values are set via mirroring? Should there be a utility to do that? We have someone who needs this mirroring (due to Jed's advice) but appear not to have any code to make that trivial. How shall we fix this? Thanks Barry

1 0

Re: [petsc-dev] [petsc-users] make test failed
by Barry Smith 20 Jun '12

20 Jun '12

Add a little python function that is called at the beginning of ./configure that goes through all the scripts and removes \r if they exist in the scripts. On non-Windows the function will do nothing and on Windows if either Mecurial or properly extracted tarball will do nothing. Will only do something if user extracted tarball wrong. Problem solved, doesn't require changing anything else in PETSc. Barry On Jun 19, 2012, at 11:51 PM, Satish Balay wrote: > On Tue, 19 Jun 2012, Barry Smith wrote: > >>> It appears that the extra '\r' char is causing bash to process strings >>> incorrectly. > >> Yes, but exactly how? Run the winzip to get the extra \r in >> them. > > BTW: unix2dos is equivalent to winzip wrt reproduction. > >> Then start removing them as you run the script to determine >> exactly where the grief comes from. I do not think the extra \r are >> always harmful, just if they come up in the wrong place. If you >> determine this you can reformat the original code to not have the >> bad \r cause grief. > > Assume the extra char is 'x' [instead of '\r'] for illustration. > > i.e: > var='foo'\r > is represented as > var='foo'x > > Note: the above is valid syntax [but 'var' will have extra stuff that > wasn't intended] > > > The primary issue is - if you have the following code [in dos formatted text]: >>>> > if [ 'a' == 'a'] > then > echo "yes" > else > echo "no" > fi > <<< > It is interpreted as: >>>> > if [ 'a' == 'a']x > thenx > echo "yes"x > elsex > echo "no"x > fix > <<< > > And you get the error "syntax error: unexpected end of file" > > One way to workarround is to add 'var=foo' before '\r' and > get some valid bash syntax. i.e something like: > >>>> > if [ 'a' == 'a'] then a='a'x > echo "yes"; a='a'x > else a='a'x > echo "no"; a='a'x > fi; a='a'x > <<<< > > So the fixed mpiexec.uni looks something like: >>>>>> > $ cat mpiexec.uni > #!/bin/sh > if [ $1 != "-np" -a $1 != "-n" ]; then a=a > progname=$*; a=a > elif [ $2 = "1" ]; then a=a > shift ; shift; a=a > progname=$* ; a=a > else a=a > echo "Uniprocessor version of MPI can only use one processor"; a=a > exit 1; a=a > fi; a=a > a=a > # Execute the program > $progname;a=a > exit 0;a=a > <<<<<<< > > There is an alternative of setting the following env before running > bash scripts in cygwin - but these commands appear to be cygwin/bash > specific [so cannot easily automate?] > >>>>>>>> > export SHELLOPTS > set -o igncr > <now run the script> > <<<<<<<< > > Satish >

1 0

Re: [petsc-dev] call stack
by John Fettig 20 Jun '12

20 Jun '12

You can use hpctoolkit to profile the cpu portion, it does exactly what you are asking. To profile the cuda portion, you will have to use nvidia's visual profiler. John On Wed, Jun 20, 2012 at 4:32 AM, Luigia Ambrosio <ambrosio.luigia(a)gmail.com> wrote: > Hi all > I'm interested in using PETSC with CUDA. I just downloaded and installed the > developer version and I was testing the example #1 > ($HOME/petsc-dev/src/vec/vec/examples/tutorials) to compute the norm of a > vector. > I would like to compare PETSC "without cuda" and "with cuda". Is there a way > to profile a PETSC executable, obtain the stack of the called functions and > methods and access to the function body? > Can anybody make an exmple? > Thanks a million

1 0

Re: [petsc-dev] call stack
by Barry Smith 20 Jun '12

20 Jun '12

On Jun 20, 2012, at 4:18 AM, Luigia Ambrosio wrote: > Yes but it's not enough! I nedd the stack of the called methods and kernels. > I just sent you my modified ex11 and the results of ./ex11 -log_summary > <ex11.c> > <log_summary> > thanks again gprof? Or the developer tools that Intel sells or Apple includes with XCode? Or Tau? Is that the kind of thing you are looking for. Apple's are fairly good for seeing the stack and where it spends time with the lines of source code etc. None of these is particular to PETSc but. Barry > > Luigia Ambrosio > > > > > Il giorno 20/giu/2012, alle ore 10:45, Aron Ahmadia ha scritto: > >> Have you tried passing -log_summary to the command line? >> >> A >> >> On Wed, Jun 20, 2012 at 11:32 AM, Luigia Ambrosio <ambrosio.luigia(a)gmail.com> wrote: >> Hi all >> I'm interested in using PETSC with CUDA. I just downloaded and installed the developer version and I was testing the example #1 ($HOME/petsc-dev/src/vec/vec/examples/tutorials) to compute the norm of a vector. >> I would like to compare PETSC "without cuda" and "with cuda". Is there a way to profile a PETSC executable, obtain the stack of the called functions and methods and access to the function body? >> Can anybody make an exmple? >> Thanks a million >> >

1 0

Re: [petsc-dev] call stack
by Matthew Knepley 20 Jun '12

20 Jun '12

On Wed, Jun 20, 2012 at 3:18 AM, Luigia Ambrosio <ambrosio.luigia(a)gmail.com>wrote: > Yes but it's not enough! I nedd the stack of the called methods and > kernels. > I just sent you my modified ex11 and the results of ./ex11 -log_summary > > > Use the debugger. Its the easiest way to get a stack. However, for the old people here, it is also possible to do by looking at the source code. Matt > thanks again > > Luigia Ambrosio > > > > > Il giorno 20/giu/2012, alle ore 10:45, Aron Ahmadia ha scritto: > > Have you tried passing -log_summary to the command line? > > A > > On Wed, Jun 20, 2012 at 11:32 AM, Luigia Ambrosio < > ambrosio.luigia(a)gmail.com> wrote: > >> Hi all >> I'm interested in using PETSC with CUDA. I just downloaded and installed >> the developer version and I was testing the example #1 >> ($HOME/petsc-dev/src/vec/vec/examples/tutorials) to compute the norm of a >> vector. >> I would like to compare PETSC "without cuda" and "with cuda". Is there a >> way to profile a PETSC executable, obtain the stack of the called functions >> and methods and access to the function body? >> Can anybody make an exmple? >> Thanks a million >> > > > > -- 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] call stack
by Luigia Ambrosio 20 Jun '12

20 Jun '12

Yes but it's not enough! I nedd the stack of the called methods and kernels. I just sent you my modified ex11 and the results of ./ex11 -log_summary thanks again Luigia Ambrosio Il giorno 20/giu/2012, alle ore 10:45, Aron Ahmadia ha scritto: > Have you tried passing -log_summary to the command line? > > A > > On Wed, Jun 20, 2012 at 11:32 AM, Luigia Ambrosio <ambrosio.luigia(a)gmail.com> wrote: > Hi all > I'm interested in using PETSC with CUDA. I just downloaded and installed the developer version and I was testing the example #1 ($HOME/petsc-dev/src/vec/vec/examples/tutorials) to compute the norm of a vector. > I would like to compare PETSC "without cuda" and "with cuda". Is there a way to profile a PETSC executable, obtain the stack of the called functions and methods and access to the function body? > Can anybody make an exmple? > Thanks a million >

1 0

Re: [petsc-dev] call stack
by Aron Ahmadia 20 Jun '12

20 Jun '12

Have you tried passing -log_summary to the command line? A On Wed, Jun 20, 2012 at 11:32 AM, Luigia Ambrosio <ambrosio.luigia(a)gmail.com > wrote: > Hi all > I'm interested in using PETSC with CUDA. I just downloaded and installed > the developer version and I was testing the example #1 > ($HOME/petsc-dev/src/vec/vec/examples/tutorials) to compute the norm of a > vector. > I would like to compare PETSC "without cuda" and "with cuda". Is there a > way to profile a PETSC executable, obtain the stack of the called functions > and methods and access to the function body? > Can anybody make an exmple? > Thanks a million >

1 0

call stack
by Luigia Ambrosio 20 Jun '12

20 Jun '12

Hi all I'm interested in using PETSC with CUDA. I just downloaded and installed the developer version and I was testing the example #1 ($HOME/petsc-dev/src/vec/vec/examples/tutorials) to compute the norm of a vector. I would like to compare PETSC "without cuda" and "with cuda". Is there a way to profile a PETSC executable, obtain the stack of the called functions and methods and access to the function body? Can anybody make an exmple? Thanks a million

1 0