h5py: Tests with long-double precision fail on some architectures

In addition to the test failures reported in #816, builds on architectures which do not support long-double precision (such as ppc64el) produce the following output:

======================================================================
FAIL: test_complex256 (h5py.tests.old.test_dataset.TestCreateShape)
Confirm that the default dtype is float
----------------------------------------------------------------------
Traceback (most recent call last):
  File "h5py/tests/old/test_dataset.py", line 101, in test_complex256
    self.assertEqual(dset.dtype, np.dtype('complex256'))
AssertionError: dtype('complex128') != dtype('complex256')

======================================================================
FAIL: test_long_double (h5py.tests.old.test_dataset.TestCreateShape)
Confirm that the default dtype is float
----------------------------------------------------------------------
Traceback (most recent call last):
  File "h5py/tests/old/test_dataset.py", line 94, in test_long_double
    self.assertEqual(dset.dtype, np.longdouble)
AssertionError: dtype('float64') != <type 'numpy.float128'>

About this issue

Original URL
State: closed
Created 7 years ago
Comments: 21 (21 by maintainers)

Most upvoted comments

https://buildd.debian.org/status/logs.php?pkg=h5py&ver=2.7.0~rc2-2 <- new logs.

It looks like on powerpc and ppc64 we are getting tests that fail due to the system (correctly?) deciding that it does not have enough bits. The ppc64el one is more worrying as there is some miss-match between the names, the actually bit depth, and what numpy thinks the bit depth is.

I would argue that only the ppc64el one should be a blocking issue (as the other ones could be dealt with by known-fails).

tacaswell on Jan 13, 2017