diff options
| author | Charles Harris <charlesr.harris@gmail.com> | 2017-07-26 20:02:30 -0600 |
|---|---|---|
| committer | Charles Harris <charlesr.harris@gmail.com> | 2017-08-03 16:38:56 -0600 |
| commit | 18a4a4b093c891c6cd89eac408a2a3214d19def1 (patch) | |
| tree | f76f530f11f5e810473c1f55d55b9be18b87d9bd /numpy/lib/tests/test_index_tricks.py | |
| parent | a646a15b86d7fd77f1b6a92e8be852cddaff41ab (diff) | |
| download | numpy-18a4a4b093c891c6cd89eac408a2a3214d19def1.tar.gz | |
BUG: Fix true_divide when dtype=np.float64 specified.
The problem here was due to the signature matching when dtype is
specified. A signature matches when
1) The output dtype matches the specified dtype
2) The input arguments can be cast to the signatures dtypes using
casting="same_kind"
The signature 'bb->d' was present for true_divide, consequently when
dtype=np.float64 was specified, all integer inputs were cast to int8.
The solution adopted here was to remove all signatures/loops for
true_divide that had integer inputs. Lesser measures broke a lot of
existing code that relied on the behavior for other functions
Closes #3484. Again.
Diffstat (limited to 'numpy/lib/tests/test_index_tricks.py')
0 files changed, 0 insertions, 0 deletions