Propagated Flags Mismatch

Comparing propagation results is showing a rare difference in the step from material. Reproduce with:

delta:~ blyth$ cd /usr/local/env/tmp/1
delta:1 blyth$ daephotonscompare.sh --loglevel debug

The npz files being compared are pulled off the GPU into numpy arrays.

Nine out of ~5000:

In [3]: np.all( a.propagated['flags'] == b.propagated['flags'] )
Out[3]: False

In [4]: np.where( a.propagated['flags'] != b.propagated['flags'] )
Out[4]:
(array([32990, 32998, 32999,
        33180, 33188, 33189,
        33460, 33468, 33469]),
 array([3, 3, 3, 3, 3, 3, 3, 3, 3]))

In [5]: indice = np.where( a.propagated['flags'] != b.propagated['flags'] )

In [6]: a.propagated['flags'][indice]
Out[6]: array([3, 3, 3, 8, 8, 8, 8, 8, 8], dtype=uint32)

In [7]: b.propagated['flags'][indice]
Out[7]: array([4, 4, 4, 4, 4, 4, 5, 5, 5], dtype=uint32)

Discrepants all have the same pattern with the start material code differing:

In [8]: a.propagated['flags'][32990:32990+10]
Out[8]:
array([[         1, 1103839338, 1103839338,          3],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         1, 1103839338, 1103839338,          3],
       [         1, 1103839338, 1103839338,          3]], dtype=uint32)

In [9]: b.propagated['flags'][32990:32990+10]
Out[9]:
array([[         1, 1103839338, 1103839338,          4],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         0,          0,          0,          0],
       [         1, 1103839338, 1103839338,          4],
       [         1, 1103839338, 1103839338,          4]], dtype=uint32)

Have seen issues of varying material codes on invokation, but this appears not to be that, as the numbers of discrepancies is low and not covering all instances of the code:

In [18]: len(np.where( a.propagated['flags'][:,3] == 3 )[0])
Out[18]: 6319

In [19]: len(np.where( b.propagated['flags'][:,3] == 3 )[0])
Out[19]: 6316

In [21]: len(np.where( a.propagated['flags'][:,3] == 4 )[0])
Out[21]: 2172

In [20]: len(np.where( b.propagated['flags'][:,3] == 4 )[0])
Out[20]: 2178

In [22]: len(np.where( a.propagated['flags'][:,3] == 5 )[0])
Out[22]: 11204

In [23]: len(np.where( b.propagated['flags'][:,3] == 5 )[0])
Out[23]: 11207