diff options
| author | Raghuveer Devulapalli <raghuveer.devulapalli@intel.com> | 2020-02-20 13:38:07 -0800 |
|---|---|---|
| committer | Raghuveer Devulapalli <raghuveer.devulapalli@intel.com> | 2020-02-20 13:47:30 -0800 |
| commit | 3253ab7c9c1f6faea0e5eb0a906db1a08ef42026 (patch) | |
| tree | 117be0dd477334a4c34edc5733c16a9bb140f28d /numpy/array_api/tests/test_array_object.py | |
| parent | 473a68a53028f8a24b88bb9fc08b38dc3b69800e (diff) | |
| download | numpy-3253ab7c9c1f6faea0e5eb0a906db1a08ef42026.tar.gz | |
BUG: Remove check requiring float/double natural alignment for AVX code
In a x86-32 bit system, doubles need not be naturally aligned to 8 Byte
boundary (see: -malign-double section of
https://gcc.gnu.org/onlinedocs/gcc/x86-Options.html). Having this check
meant it ran different code paths (AVX v/s scalar) depending on the
alignment of data which leads to different results and test failing
intermittently. AVX code uses un-aligned loads and this check is
unnecessary to begin with.
Diffstat (limited to 'numpy/array_api/tests/test_array_object.py')
0 files changed, 0 insertions, 0 deletions