v2.2.0 - Adjoint flow and cleaning of tbox

Waves v2.2.0 (flow v1.8.0) - Adjoint solver

This release adds a first version of a validated adjoint solver for flow. It enables the computation of gradients with respect to lots of design variables (such as the mesh coordinates), for a low cost. Additionally, the way of building element matrices for the different modules have been reviewed, and several classes have been refactored or cleaned accordingly.

Changes

• Add adjoint flow solver for gradients computation
• Clean Element/Mem classes (#49)
• Refactor functions
• Minor changes and bugfixing (thx @R.Boman)

Notes

Element/Mem

• All the build() methods that were specific to a given physics have been removed from the Element class. These methods have been move to their respective module in ...Term, ...Residual and ...Functional classes. Additionally, the methods are now static.
• Element now only stores geometric and basic computational information. Accordingly, all the data previously stored in the Mem class gave been gathered into Element, and Mem has been removed. The data (Jacobian, normal vector, etc.) must be initialized before any computation can be carried out. This action is performed using the initValues() method. Additionally, the order of integration for the Gaussian quadrature can be set using this method. Since the gradients of some data (Jacobian, etc.) is required for some computations, they can be initialized using a separate method initGradients(). Once all the data have been initialized, the can be updated as before using the update() method.

Functions

• All the functions used in flow (F0El, etc.) are now specific to this module. Additionally, they derive from the new Observer class and are not wrapped in python anymore.
• An Observer class has been added to fwk. It provides a mechanism for self-updating, based on changes occurring in a given class.
• The functions of tbox (Fct0, etc.) are now smart pointers (40171e0c (comment 5682) has been fixed). Note that these functions are only used in heat and could be moved outside of tbox.

Tests

Tests are passing on ubuntu20.04 (python 3.8.2), and msys2 (python 3.8.1).

This commit adds smart pointer support for the function classes in tbox (Fct...), heat (CompiledFct...) and flow (F...).

At this point the code is:

• compiling and running in Release config
• compiling and running in RelWithDebugInfo config, except for flow tests
• compiling but failing at runtime in Debug config, for all tests

I could not identify the cause of the failure at runtime in debug, but this happens since v1.7.1 at least. So, this commit is not responsible for that failure and I will not talk about it.

Regarding the failure at runtime for flow tests, it is triggered by these lines:

initial = Initial(msh, "group", F0PsPhiInf(...))
dirichlet = Dirichlet(msh, "group", F0PsPhiInf(...))

where Initial and Dirichlet both derive from Assign, and F0PsPhiInf derives from F0Ps. The constructor of Assign is Assign(shared_ptr<MshData>, string, shared_ptr<F0Ps>).
More specifically, the cause of the failure is an assertion triggered in a SWIG generated function (that explains why it does not fail in release):

int SWIG_Python_ConvertPtrAndOwn(PyObject*, void**, swig_type_info*, int, int*) {
// ...
if (own)
{
  int newmemory = 0;
  *ptr = SWIG_TypeCast(tc,vptr,&newmemory);
  if (newmemory == SWIG_CAST_NEW_MEMORY)
  {
    assert(own); /* badly formed typemap which will lead to a memory leak - it must set and use own to delete *ptr */
      if (own)
         *own = *own | SWIG_CAST_NEW_MEMORY;
  }
}
// ...
}

where int *own is the last parameter passed to SWIG_Python_ConvertPtrAndOwn.
After a few tests, I realized that printing *own resulted in a segfault. Note that several of my classes pass through this bit of code, but never trigger the assertion since *own = 0. I do not understand what is happening there.

What is even more puzzling is that I have other classes, which do exactly the same thing, but with function classes from tbox, and they do not fail. A simple example is the Freestream(shared_ptr<MshData>, string, shared_ptr<Fct1>) class, that I instantiate in python as freestream = Freestream(msh, "group", Fct1C(...)), where Fct1C derives from Fct1.

Finally, I should mention that I ran valgrind on both the RelWithDebugInfo (after commenting out the assert(own)) and the Release configs, and that it detected no memory leaks.

In the end, I have several questions:

1. What is happening and why?
2. Why do the classes deriving from flow::F0Ps cause an issue, while the classes deriving from tbox::Fct0 do not?
3. Can we trust valgrind when it says that there is no memory leak (knowing that the comment after the assertion in the SWIG function indicates that there will be a memory leak)?

Additional notes if you got so far without dying:

1. This issue https://github.com/swig/swig/issues/731 is still opened in the SWIG repo and is basically the same I run through
2. If I change Assign to accept the derived class F0PsPhiInf (instead of the base F0Ps), it works fine. I also tried to replace F0ps by Fct0, and it runs.
3. I encountered no issue, nor memory leaks due to the combined use of shared_ptr and director, which was my initial fear.

Many thanks in advance.

On ubuntu 18.04, all tests are now running whether in Release or Debug.

• mesh size has been slightly adjusted in nonlift test case to ensure proper convergence
• Dirichlet BCs have been kept in bli test case due to lack of robustness with all Neumann BCs, they will have to be removed at some point
• the solver is slightly slower (2%) when only Neumann BCs are used (consistent, because it takes more time to compute and assemble a Neumann BC than a Dirichlet BC)

Also added the correct derivative for the limited density

Note that the variable order is not yet accessible outside element and defaults to 2. A setter method will be added when refactoring Element/Mem classes (#49).

added cleaning enhancement labels

The build works on Windows with Visual Studio 2017.

Only lift3 fails:

98% tests passed, 1 tests failed out of 63

The following tests FAILED:
         22 - flow/tests/lift3.py (Failed)
Errors while running CTest

with the following results:

[CTest] iteration count = 3.000000 (expected 3.000000 +/- 1.000000)
	abs diff = 0.000000e+00 <= 1.000000e+00 [ok]
[CTest] CL = 0.132850 (expected 0.135000 +/- 5.0%)
	rel diff = 1.592774e-02 <= 5.000000e-02 [ok]
[CTest] CD = 0.006489 (expected 0.006200 +/- 0.010000)
	abs diff = 2.886878e-04 <= 1.000000e-02 [ok]
[CTest] CS = 0.012392 (expected 0.012000 +/- 1.0%)
	rel diff = 3.268972e-02 > 1.000000e-02 [wrong!]
[CTest] CM = -0.027948 (expected -0.027800 +/- 1.0%)
	rel diff = 5.319034e-03 <= 1.000000e-02 [ok]

added 1 commit

• 860b8e05 - fix test for msvs2017

Compare with previous version

marked this merge request as ready

changed title from Draft: 2.2.0 - Adjoint flow and cleaning of tbox to v2.2.0 - Adjoint flow and cleaning of tbox

changed the description

Ok, everything works with visual studio 2017

100% tests passed, 0 tests failed out of 63
Total Test time (real) =  30.87 sec

approved this merge request

merged