cudnnTransformTensor & cudnnAddTensor interpret stride different

This is now puzzling me for for 3 hours. My expectation would be that the results of the following two operations should be that same. However, they are not. Note that addTensor seems to do what I would expect. Am I getting this correct? According to the manual I should prefer transformTensor over addTensor for reshaping. Why is transformTensor behaving differently here?

alpha = 1.0f

beta = 0.0f

dataPtr: dims=[2, 1, 2, 1], stride=[2, 2, 1, 1], content=[1.0f, 2.0f, 3.0f, 4.0f], format=FLOAT]

dstDataPtr: dims=[2, 1, 2, 1], stride=[6, 2, 1, 1], content=[NaN, NaN, …, NaN ], format=FLOAT]

cudnnTransformTensor(
cudnnPtr,
&alpha, descPtr, dataPtr,
&beta, dstDescPtr, dstdataPtr
)

RESULT:
dstDescPtr = [1.0f, 2.0f, 3.0f, 4.0f, NaN, NaN, NaN, NaN, NaN, NaN, NaN, NaN]

cudnnAddTensor(
cudnnPtr,
&alpha, descPtr, dataPtr,
&beta, dstDescPtr, dstdataPtr
)

RESULT:
dstDescPtr = [1.0f, 2.0f, NaN, NaN, NaN, NaN, 3.0f, 4.0f, NaN, NaN, NaN, NaN]