deal.II with PETSc: compiling with support for complex numbers

[Artur Safin]

Hello all,

I have an installation of PETSc with support for complex numbers ('--with-scalar-type=complex' as the installation argument), and now I am trying to compile deal with this installation of PETSc. I am getting errors that look like

-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

/home/artur/dealii-8.3.0/source/numerics/data_out_dof_data.cc:In instantiation of ‘void dealii::internal::DataOut::DataEntry<DH, VectorType>::get_function_values(const dealii::FEValuesBase<DH:: dimension, DH:: space_dimension>&, std::vector<double>&) const [with DH = dealii::hp::DoFHandler<1, 3>; VectorType = dealii::PETScWrappers::MPI::BlockVector]’:

/home/artur/dealii-8.3.0/source/numerics/data_out_dof_data.cc:1251:1:   required from here

/home/artur/dealii-8.3.0/source/numerics/data_out_dof_data.cc:589:25: error: cannot convert ‘__gnu_cxx::__alloc_traits<std::allocator<std::complex<double> > >::value_type {aka std::complex<double>}’ to ‘__gnu_cxx::__alloc_traits<std::allocator<double> >::value_type {aka double}’ in assignment

         patch_values[i] = tmp[i];

                         ^

-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------

For a standard PETSc installation (without support for complex numbers), deal compiles without any problems. I have two questions:

1) Does deal support complex PETSc vectors and matrices?

2) Is there a way to work around the error above?

Best,

Artur

[Timo Heister]

> 1) Does deal support complex PETSc vectors and matrices?

While we honor the PetscScalar type in the wrapper classes, this is
not used/tested very well and I wouldn't be surprised if you run into
problems.

> 2) Is there a way to work around the error above?

This is related to complex support inside deal.II, which is work in
progress. Please see
https://github.com/dealii/dealii/wiki/Frequently-Asked-Questions#can-i-solve-problems-over-complex-numbers
to avoid matrices/vectors with complex entries.

--
Timo Heister
http://www.math.clemson.edu/~heister/

[Denis Davydov]

Hi Arthur,

yes, but not on the master. There are still a couple of places which won't compile.

The most problematic is the DataOutput class.

 

2) Is there a way to work around the error above?

yes, you can do it in deal.II even without doing the split into real and imaginary part.

see this branch https://github.com/davydden/dealii/tree/branch_petscscalar_complex

Alternatively, and actually preferably, would be if you could give the current master a look and

for starters try disabling instantiation of DataOut class for PETSc vectors in case

of complex-valued PETSc. It very-well could be the only not-coverd part of deal.II yet.

Then the work-around would be to copy PETSc vectors into, say, Trilinos for output only.

If you are lucky all the weak forms and solvers would stay with complex algebra.

Regards,

Denis.

[Wolfgang Bangerth]

> 2) Is there a way to work around the error above?
>
> yes, you can do it in deal.II even without doing the split into real and
> imaginary part.
> see this branch https://github.com/davydden/dealii/tree/branch_petscscalar_complex

I'm willing to accept incremental patches that get master in the direction you
want to go -- it doesn't have to be the one grand patch that does it all in
one fall swoop :-)

Cheers
W.
--
------------------------------------------------------------------------
Wolfgang Bangerth email: bang...@math.tamu.edu
www: http://www.math.tamu.edu/~bangerth/

[Denis Davydov]

On Monday, November 16, 2015 at 12:05:39 AM UTC+1, bangerth wrote:

>     2) Is there a way to work around the error above?
>
> yes, you can do it in deal.II even without doing the split into real and
> imaginary part.
> see this branch https://github.com/davydden/dealii/tree/branch_petscscalar_complex

I'm willing to accept incremental patches that get master in the direction you
want to go -- it doesn't have to be the one grand patch that does it all in
one fall swoop :-) 

We already have most of that branch in master (e.g. FEValues, constraint matrix, etc).

What is left:

1. vector and scalar list for complex/real valued

https://github.com/davydden/dealii/blob/branch_petscscalar_complex/cmake/config/template-arguments.in#L2-L5

https://github.com/davydden/dealii/blob/branch_petscscalar_complex/cmake/config/template-arguments.in#L68-L73

https://github.com/davydden/dealii/blob/branch_petscscalar_complex/cmake/config/template-arguments.in#L68-L73

2. use of thereof to mask some parts of the library which are not yet extended to complex numbers (DataOut is a pandora box!..)

Now thinking about it, the alternative could be to provide explicit specialisation of un-supported classes for complex vectors where needed

when PETSc is complex. Theoretically that should not require changing/introducing new  definitions.

Regards,

Denis.

[Artur Safin]

Timo, Denis,

Thank you for your suggestions. 

I am aware that you can avoid these problems by doubling the basis, but the problem I am solving may perform better with a complex numbers formulation, so it makes sense to go this route.

Denis, I have tried to compile the master against complex PETSc. After getting past errors in data_out, I ran into similar issues in the derivative_approximation & error_estimator classes - I stopped after that, since at this point I was commenting out essential parts of the package. I'll try to see if I can get your branch to work.

Best,

Artur

[Wolfgang Bangerth]

> What is left:
> 1. vector and scalar list for complex/real valued
> https://github.com/davydden/dealii/blob/branch_petscscalar_complex/cmake/config/template-arguments.in#L2-L5
> https://github.com/davydden/dealii/blob/branch_petscscalar_complex/cmake/config/template-arguments.in#L68-L73
> https://github.com/davydden/dealii/blob/branch_petscscalar_complex/cmake/config/template-arguments.in#L68-L73

This should be uncontroversial. Feel free to get that off your branch and onto
mainline with a simple patch.

> 2. use of thereof to mask some parts of the library which are not yet extended
> to complex numbers (DataOut is a pandora box!..)

I don't think it's actually going to be that difficult. See
https://github.com/dealii/dealii/issues/1894
Or do you already have a different implementation?

Best

[Denis Davydov]

FYI, i made the master branch compile with complex-valued PETSc (with some minor masking of parts of the library described here: https://github.com/dealii/dealii/issues/2033#issuecomment-187159252), 

there are  a number of pull-reqests pending (aggregated in this issue https://github.com/dealii/dealii/issues/2033 ).

So it will be possible to use the up-to-date deal.II in that scenario very soon...

Regards,

Denis.