[petsc-users] How to interface with PETSc from another application?

[Jed Brown]

Klaus Burkart <k_burkart at yahoo.com> writes:

> When I link the petsc library, the application side code is not properly compiled and the solver is not available for selection
>
> --> FOAM FATAL IO ERROR:
> Unknown asymmetric matrix solver petGMRES

There must be some earlier error message.

> Valid asymmetric matrix solvers are :
>
> 4
> (
> GAMG
> PBiCG
> PBiCGStab
> smoothSolver
> )
> I added the following to my makefile to link the petsc library
>
>     -L$(PETSC_DIR)/arch-linux2-c-debug/lib    -lpetsc

In this case, you'd need to add that path to LD_LIBRARY_PATH so the
loader can find it.  None of these are PETSc issues, just linking
dynamic libraries.

>     Jed Brown <jed at jedbrown.org> schrieb am 18:45 Mittwoch, 4.Oktober 2017:
>  
>
>  Klaus Burkart <k_burkart at yahoo.com> writes:
>
>> My setup:
>>
>> .bashrc
>> export PETSC_DIR=/home/klaus/OpenFOAM/klaus-5.0/petsc-3.7.6
>> export PETSC_ARCH=arch-linux2-c-debug
>> export PETSC_CONFIGDIR=${PETSC_DIR}/lib/petsc
>>
>>
>> make options
>>     -I$(PETSC_CONFIGDIR)/conf \
>
> The above should not be needed.
>
>>     -I$(PETSC_DIR)/include \
>>     -I$(PETSC_DIR)/arch-linux2-c-debug/include
>>
>> Installation and tests worked fine.
>>
>> The output using:     PetscInitialize(0,0,NULL,NULL); at the beginning and PetscFinalize(); at the end of the code section including PETSc (solver section) is:
>>
>> simpleFoam: symbol lookup error: /home/klaus/OpenFOAM/klaus-5.0/platforms/linux64GccDPInt32Opt/lib/libpetFoam.so: undefined symbol: PetscInitialize
>
> How have you linked libpetFoam.so to libpetsc.so and did you use RPATH?
> That seems to be the problem.
>
>> No simulation is triggered
>>
>>
>> When I just declare Mat M; and call a function with M as a parameter which outputs "Petsc - Hello" and sets the matrix M to symmetric (using: MatSetOption(A, MAT_SYMMETRIC, PETSC_TRUE);), the execution of a simulation is triggered but MatSetOption is causing a problem, I assume because PetscInitialize is missing
>>
>> Petsc - Hello nonepetGMRES:  Solving for Ux, Initial residual = 1, Final residual = 1, No Iterations 0
>> Petsc - Hello nonepetGMRES:  Solving for Uz, Initial residual = 1, Final residual = 1, No Iterations 0
>> simpleFoam: symbol lookup error: /home/klaus/OpenFOAM/klaus-5.0/platforms/linux64GccDPInt32Opt/lib/libpetFoam.so: undefined symbol: MatSetOption
>>
>> Maybe important to know, there's no way to enter commmand line input in the terminal while a simulation is running because the application displays continuously the intermediate simulation results. That's why I use PetscInitialize(0,0,NULL,NULL); There's now way to provide command line input.
>
> That's fine.  You can use the PETSC_OPTIONS environment variable or a
> configuration file to get run-time options to PETSc.
>
>> I came back to this simple test after writing "the complete code" which showed these problems and stripped it down, step-by-step, to figure out what causes the problem i.e. everything but a declaration. 
>>
>>  
>>
>>    Jed Brown <jed at jedbrown.org> schrieb am 17:17 Mittwoch, 4.Oktober 2017:
>>  
>>
>>  Klaus Burkart <k_burkart at yahoo.com> writes:
>>
>>> What's the concept to interface with PETSc from another application to solve a linear system with PETSc?
>>>
>>> The standard procedure of the job:
>>>
>>> 1: The application provides a matrix which needs to be converted and be loaded into PETSc
>>>
>>> 2: The application provides the rhs vector (containing pointers!) which needs to be loaded into PETSc 
>>>
>>> 3: The linear system is to be solved using PETSc
>>>
>>> 4: The application provides the result vector x, the PETSc result needs to be copied back to the application into vector x (also expecting pointers)
>>>
>>>
>>> The problem - maybe a completely wrong approach when it comes to using PETSc: 
>>>
>>> With other linear algebra libraries, I included the library functionality in the code of a new solver accessing the functionality usually via header files and created a plugin which can be called from the application when running a simulation. 
>>>
>>> Even so the mixed code including PETSc code can be compiled, the bit of the plugin, interfacing with the application is broken as soon as I include more than a PETSc declaration in the mixed code.
>>
>> Sounds like maybe you haven't correctly linked to the PETSc library.
>> Sending us the commands run and output/errors would be helpful to debug.
>>
>>> How to interface with PETSc from a software application?  (I am using c++ and Ubuntu)
>>>
>>> Klaus
>>
>>    
>
>