MDagPath.inclusiveMatrix() doesn't give reliable/deterministic results

[danny269DZ]

I have a problem where iterating over objects and calling MDagPath.inbclusiveMatrix() doesn't give deterministic results. Specifically, if I:

1. Generate or load a scene
2. Iterate over all transform objects and dump out their attributes and their world-space matrix obtained from inclusiveMatrix (script below)
3. Save the scene in binary format
4. Make a new empty scene (just to flush things)
5. Load the scene
6. Iterate and print attributes and matrices again

then comparing the output of steps 2 and 6 (i.e. before and after the save/load):

1. The attributes match up perfectly (to 64 sig digits)
2. Matrices returned from inclusiveMatrix differ slightly - after around 16 sig digits

This might not sound like a lot - but when these are passed to a physics solver (e.g. to initialise objects or control them), it means that a scene you've set up and tweaked will sometimes behave slightly/significantly/wildly differently after a save/load operation.

Has anybody got any ideas what might cause these differences? And even better - any idea how I might either flush the system so that inclusiveMatrix is forced to discard any out-of-date data it might be using, or get the world-space matrices a different (more reliable) way? Calling OpenMaya.MFnMatrixData on the "worldMatrix" plug results in the same data as inclusiveMatrix, unfortunately.

Tested with Maya 2018 and 2020 under Windows (though I believe it happens under Linux too)

Thanks for any ideas!

Note: I lied slightly when I said the attributes match up perfectly: "-0" tends to turn into "0". However, this doesn't appear to be related to the times that inclusiveMatrix generates different output, so I'm hoping it isn't itself a problem.

Here's my script to dump the attributes and world matrix:

from maya import cmds, OpenMaya

transforms = cmds.ls(type="transform")

for t in transforms:
    slls = OpenMaya.MSelectionList()
    slls.add(t)
    path = OpenMaya.MDagPath()
    slls.getDagPath(0, path)
    mat = path.inclusiveMatrix()
    for i in range(0, 4):
        print(t, 
            "{:.64g}".format(mat(0, i)),
            "{:.64g}".format(mat(1, i)),
            "{:.64g}".format(mat(2, i)),
            "{:.64g}".format(mat(3, i)))
            
for t in transforms:
    x = cmds.getAttr(t + ".translateX")
    y = cmds.getAttr(t + ".translateY")
    z = cmds.getAttr(t + ".translateZ")    
    rx = cmds.getAttr(t + ".rotateX")
    ry = cmds.getAttr(t + ".rotateY")
    rz = cmds.getAttr(t + ".rotateZ")    
    print(t, "t {:.64g}".format(x), "{:.64g}".format(y), "{:.64g}".format(z))
    print(t, "r {:.64g}".format(rx), "{:.64g}".format(ry), "{:.64g}".format(rz))

The most reliable test (because not all objects suffer from this) is to:

1. Add the HumanIK animation example to an empty scene
2. Run the script above and save the output to before.txt
3. save as binary. new file. load the saved file
4. Run the script again and save the output after.txt
5. File compare before.txt and after.txt

It also manifests with simple objects - e.g. a few poly objects under some transforms with non-identity rotation/translation/scaling.

[danny269DZ]

I still don't have a solution/explanation for this, but I do have an extremely simple reproducible example of it - demonstrated here:

https://youtu.be/2M4iCrSdL9k

The video description contains a link to the test scene, plus the script to dump out the attributes. The basic steps are:

* Load test7.ma

* Dump out attributes to before.txt

* Save scene as test7.mb

* Load scene test7.mb

* Dump out attributes to after.txt

* Compare and note that jointOrients are not preserved

It seems that saving/loading in Maya binary format preserves translate and rotate attributes on joints perfectly, but NOT jointOrient attributes.

Any ideas about this would be much appreciated!