JIT can fail silently - seems to happen when there is a problem in dolfin.pc

Can you provide an example? It's hard to debug otherwise and it's blocking 1.0-rc1.

On 20 November 2011 18:03, Anders Logg <email address hidden> wrote:
> Can you provide an example?

No, which is the problem. I get it on a particular machine, and it
seems to be related to when I link libraries, like PETSc and Trilinos,
to shared libs which are not in default locations. C++ is fine.

Do we have a test for JIT that does not involve FFC? I suspect that
our rubbish dolfin.pc is part of the problem.

Garth

> It's hard to debug otherwise and it's
> blocking 1.0-rc1.
>
> --
> You received this bug notification because you are a member of DOLFIN
> Core Team, which is subscribed to DOLFIN.
> https://bugs.launchpad.net/bugs/892714
>
> Title:
>  JIT can fail silently - seems to happen when there is a problem in
>  dolfin.pc
>
> To manage notifications about this bug go to:
> https://bugs.launchpad.net/dolfin/+bug/892714/+subscriptions
>

--
Garth N. Wells
Department of Engineering, University of Cambridge
http://www.eng.cam.ac.uk/~gnw20

On Sun, Nov 20, 2011 at 07:06:31PM -0000, Garth Wells wrote:
> On 20 November 2011 18:03, Anders Logg <email address hidden> wrote:
> > Can you provide an example?
>
> No, which is the problem. I get it on a particular machine, and it
> seems to be related to when I link libraries, like PETSc and Trilinos,
> to shared libs which are not in default locations. C++ is fine.
>
> Do we have a test for JIT that does not involve FFC? I suspect that
> our rubbish dolfin.pc is part of the problem.

You can either JIT-compile an Expression which does not involve FFC,
or you can set the form compiler to SFC and see if that changes
anything.

--
Anders

On 20 November 2011 20:00, Anders Logg <email address hidden> wrote:
> On Sun, Nov 20, 2011 at 07:06:31PM -0000, Garth Wells wrote:
>> On 20 November 2011 18:03, Anders Logg <email address hidden> wrote:
>> > Can you provide an example?
>>
>> No, which is the problem. I get it on a particular machine, and it
>> seems to be related to when I link libraries, like PETSc and Trilinos,
>> to shared libs which are not in default locations. C++ is fine.
>>
>> Do we have a test for JIT that does not involve FFC? I suspect that
>> our rubbish dolfin.pc is part of the problem.
>
> You can either JIT-compile an Expression which does not involve FFC,
> or you can set the form compiler to SFC and see if that changes
> anything.
>

Expressions seem to be fine. It's going to require some deep digging.

Garth

> --
> Anders
>
> --
> You received this bug notification because you are a member of DOLFIN
> Core Team, which is subscribed to DOLFIN.
> https://bugs.launchpad.net/bugs/892714
>
> Title:
>  JIT can fail silently - seems to happen when there is a problem in
>  dolfin.pc
>
> To manage notifications about this bug go to:
> https://bugs.launchpad.net/dolfin/+bug/892714/+subscriptions
>

--
Garth N. Wells
Department of Engineering, University of Cambridge
http://www.eng.cam.ac.uk/~gnw20

On 20 November 2011 20:37, Garth N. Wells <email address hidden> wrote:
> On 20 November 2011 20:00, Anders Logg <email address hidden> wrote:
>> On Sun, Nov 20, 2011 at 07:06:31PM -0000, Garth Wells wrote:
>>> On 20 November 2011 18:03, Anders Logg <email address hidden> wrote:
>>> > Can you provide an example?
>>>
>>> No, which is the problem. I get it on a particular machine, and it
>>> seems to be related to when I link libraries, like PETSc and Trilinos,
>>> to shared libs which are not in default locations. C++ is fine.
>>>
>>> Do we have a test for JIT that does not involve FFC? I suspect that
>>> our rubbish dolfin.pc is part of the problem.
>>
>> You can either JIT-compile an Expression which does not involve FFC,
>> or you can set the form compiler to SFC and see if that changes
>> anything.
>>
>
> Expressions seem to be fine. It's going to require some deep digging.
>

After a of digging, it's numpy that fails silently. It was failing
when FIAT tried to solve a linear system. I can reproduce the failure
with

    from dolfin import *
    import numpy
    A = numpy.array([[3,1], [1,2]])
    b = numpy.array([9,8])
    x = numpy.linalg.solve(A, b)
    print x

but it works if I switch the import order,

    import numpy
    from dolfin import *
    A = numpy.array([[3,1], [1,2]])
    b = numpy.array([9,8])
    x = numpy.linalg.solve(A, b)
    print x

I suspect that it's related to numpy and DOLFIN linking to different
BLAS libraries.

Garth

> Garth
>
>> --
>> Anders
>>
>> --
>> You received this bug notification because you are a member of DOLFIN
>> Core Team, which is subscribed to DOLFIN.
>> https://bugs.launchpad.net/bugs/892714
>>
>> Title:
>>  JIT can fail silently - seems to happen when there is a problem in
>>  dolfin.pc
>>
>> To manage notifications about this bug go to:
>> https://bugs.launchpad.net/dolfin/+bug/892714/+subscriptions
>>

On Mon, Nov 21, 2011 at 09:19:53AM -0000, Garth Wells wrote:
> On 20 November 2011 20:37, Garth N. Wells <email address hidden> wrote:
> > On 20 November 2011 20:00, Anders Logg <email address hidden> wrote:
> >> On Sun, Nov 20, 2011 at 07:06:31PM -0000, Garth Wells wrote:
> >>> On 20 November 2011 18:03, Anders Logg <email address hidden> wrote:
> >>> > Can you provide an example?
> >>>
> >>> No, which is the problem. I get it on a particular machine, and it
> >>> seems to be related to when I link libraries, like PETSc and Trilinos,
> >>> to shared libs which are not in default locations. C++ is fine.
> >>>
> >>> Do we have a test for JIT that does not involve FFC? I suspect that
> >>> our rubbish dolfin.pc is part of the problem.
> >>
> >> You can either JIT-compile an Expression which does not involve FFC,
> >> or you can set the form compiler to SFC and see if that changes
> >> anything.
> >>
> >
> > Expressions seem to be fine. It's going to require some deep digging.
> >
>
> After a of digging, it's numpy that fails silently. It was failing
> when FIAT tried to solve a linear system. I can reproduce the failure
> with
>
> from dolfin import *
> import numpy
> A = numpy.array([[3,1], [1,2]])
> b = numpy.array([9,8])
> x = numpy.linalg.solve(A, b)
> print x
>
> but it works if I switch the import order,
>
> import numpy
> from dolfin import *
> A = numpy.array([[3,1], [1,2]])
> b = numpy.array([9,8])
> x = numpy.linalg.solve(A, b)
> print x
>
> I suspect that it's related to numpy and DOLFIN linking to different
> BLAS libraries.

Both work for me. If you were a regular user, someone would say that
you have messed up your installation and mark this bug as invalid. ;-)

--
Anders

> Garth
>
> > Garth
> >
> >>
> >>
> >> Title:
> >>  JIT can fail silently - seems to happen when there is a problem in
> >>  dolfin.pc
> >>
> >> To manage notifications about this bug go to:
> >> https://bugs.launchpad.net/dolfin/+bug/892714/+subscriptions
> >>
>

Both work for me as well.

On 21 November 2011 11:43, Martin Sandve Alnæs
<email address hidden> wrote:
> Both work for me as well.
>

It's related to using optimised BLAS rather than the slow versions
that are part of Ubuntu, which is why it's unlikely to surface unless
linking to libraries in non-default locations.

I think that resolving this will require a good understanding of how
Python modules link to underlying libraries.

Garth

> --
> You received this bug notification because you are a member of DOLFIN
> Core Team, which is subscribed to DOLFIN.
> https://bugs.launchpad.net/bugs/892714
>
> Title:
>  JIT can fail silently - seems to happen when there is a problem in
>  dolfin.pc
>
> To manage notifications about this bug go to:
> https://bugs.launchpad.net/dolfin/+bug/892714/+subscriptions
>

Turns out that this is a bug in ATLAS 3.9.x when ATLAS builds LAPACK. dgels function segfaults, which leads to a silent failure.