Maybe the @epernod 's expert 👀 would be useful here :)

Hi,

could you develop what you mean by: every call after that does not
just give me an example of what is not working vs expected behavior.

Regarding removing of points, the process of renumbering is the same as I explained for removing tetrahedron in this issue: sofa-framework/sofa#3156

Case A:

1. I call the function to add 10 points -> 10 points are added (10 more visualized points pop up)
2. I call the function to add 10 points again -> No additional visual points pop up. If the MO is already open in the GUI, the nbPoints increases, but I cannot see them in the state. If the MO is NOT already open in the GUI, and I try to open it -> segfault

So for 2. I would expect 10 more points to show up :D

Case B:

1. I call the function to add 10 points -> 10 points are added (10 more visualized points pop up)
   [0, 1, ...., 11]
2. I call the function to delete index 0 -> The left most point disappears
3. I call the function to delete index 0 -> The right most point disappears
4. I call the function to delete index 0 -> The right most point disappears
5. I call the function to add 10 points -> Only one point is added (the most left one, that disappeared in 2.)

Which is particularly weird, because I "reposition" them after trying to add them

            with self.state.position.writeable() as state:
                for i in range(len(state)):
                    state[i] = np.array([i, 0, 0])

Still WIP ?

Still WIP ?

Yes, because I still do not know where the observed behavior comes from.

Ok I have further investigated the problem a bit, and I have a feeling that the changes are not properly passed to the MechanicalObject.
If I add the 10 points, the size of position.writeable() correctly changes to 12. If I then add or remove further points, the size remains at 12.

During the sofadevmeeting we look at this PR. any update is it still in WIP ?

During the sofadevmeeting we look at this PR. any update is it still in WIP ?

Hi @damienmarchal
There is still the bug with the incorrectly updated MechanicalObject.
Sadly I know too little about the event mechanism for topology changes to start investigating the problem properly.

During the sofadevmeeting we look at this PR. any update is it still in WIP ?

Hi @damienmarchal There is still the bug with the incorrectly updated MechanicalObject. Sadly I know too little about the event mechanism for topology changes to start investigating the problem properly.

Hello @damienmarchal,
Just a quick nod to this topic again, if you have some time.
Thanks !

Hi all,

After a quick investigation with @younesssss i would say the problem is in the c++ part. And @epernod is clearly the one to ask for insight, on my side I have zero knowledge on the topology change.

Hi @damienmarchal @hugtalbot ,
together with @epernod we found the underlying issue.

When trying to write into MechanicalObject.position, with writeable() or __setitem__, SofaPython3 checks if
the data is dirty or not in memcache.
For some reason, the d->isDirty() check always evaluates to false.
So after calling the function the first time, d is in memcache and the not-updated version of the array is used.
This is why len(MechanicalObject.position.writeable() is incorrect.

https://github.com/sofa-framework/SofaPython3/blob/master/Plugin/src/SofaPython3/DataHelper.cpp#L430-L445

py::array getPythonArrayFor(BaseData* d)
{
    auto& memcache = getObjectCache();
    if(d->isDirty() || memcache.find(d) == memcache.end())
    {
        auto capsule = py::capsule(new Base::SPtr(d->getOwner()), [](void*p){ delete static_cast<Base::SPtr*>(p); } );

        py::buffer_info ninfo = toBufferInfo(*d);
        py::array a(pybind11::dtype(ninfo), ninfo.shape,
                    ninfo.strides, ninfo.ptr, capsule);

        memcache[d] = a;
        return a;
    }
    return memcache[d];
}

Since the binding itself is functional, I would like to merge it, and then open another issue to address the actual problem.

• Long term, by making sure that d->isDirty() correctly evaluates to true.
• Short term, by checking if the shapes of d and memcache[d] are the same.

Cheers,
Paul

Hi @damienmarchal could you have you 🐍 python eyes 👁️ on above Paul's comment?

Hello @damienmarchal,
Did you push the latest changes you made on this thread? Thanks!