einsum error for mixed subscripts on s390x #12689
Description
Reproducing code example:
import numpy as np
tensor = np.random.rand(10, 10, 10, 10)
np.einsum('ijij->', tensor)Error message:
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
File "/usr/lib/python3/dist-packages/numpy/core/einsumfunc.py", line 1346, in einsum
return c_einsum(*operands, **kwargs)
ValueError: dimensions in operand 0 for collapsing index 'þ' don't match (19228356 != 10)
Numpy/Python version information:
Debian unstable on s390x
1.16.0rc2 3.7.2 (default, Jan 3 2019, 02:55:40)
[GCC 8.2.0]
I could not reproduce this on Debian stable (numpy 1.12.1), but 1.15 seems to be affected as well.
So if I add print(tensor[0][0][0]) before the np.einsum it runs fine:
>>> import numpy as np
>>> tensor = np.random.rand(10, 10, 10, 10)
>>> print(tensor[0][0][0])
[0.56143868 0.22681118 0.47354192 0.27676805 0.09566204 0.63877983
0.4243805 0.91690097 0.2695863 0.28861963]
>>> np.einsum('ijij->', tensor)
51.05293277361482
Also see psi4/psi4#1465 where this bug originated for some more debugging output that might be helpful.
Note that on s390x, a char is unsigned char unless specified otherwise, contrary to x86 architectures.