BUG: Unpickled void scalars should be contiguous #7141
Conversation
Void scalars are both C- and Fortran-contiguous, so pickling and unpickling them should result in a new void scalar that also has these flags set. Fixes numpy#7140.
| @@ -799,7 +799,7 @@ PyArray_Scalar(void *data, PyArray_Descr *descr, PyObject *base) | |||
| Py_INCREF(descr); | |||
| vobj->obval = NULL; | |||
| Py_SIZE(vobj) = itemsize; | |||
| vobj->flags = NPY_ARRAY_BEHAVED | NPY_ARRAY_OWNDATA; | |||
| vobj->flags = NPY_ARRAY_CARRAY | NPY_ARRAY_F_CONTIGUOUS | NPY_ARRAY_OWNDATA; | |||
There was a problem hiding this comment.
what is the reason you removed NPY_ARRAY_BEHAVED? (aligned and writeable)
There was a problem hiding this comment.
NPY_ARRAY_CARRAY is C-contiguous and behaved: https://github.com/numpy/numpy/blob/master/numpy/core/include/numpy/ndarraytypes.h#L861
There was a problem hiding this comment.
OK, can you please make sure there is a test for this as well though, or add one if there is not?
There was a problem hiding this comment.
I can - I will look at this now.
There was a problem hiding this comment.
Removing the aligned and writeable flags didn't cause any test failures, so I've added checks for both of them to the test_pickle_3 test as well.
There was a problem hiding this comment.
Some alignment features only matter on Sun hardware. I don't know if that affects this or not.
There was a problem hiding this comment.
I don't have access to any Sun hardware (I presume SPARC machines?) but I think I've avoided changing the aligned flag here - is there anything else I can look into related to this?
There was a problem hiding this comment.
that should't be necessary, alignment should be handled correctly and if not well deal with it when it causes a problem
most platforms don't care about alignment anyway, at worst you have software handling on those and some slowdown which is irrelevant for scalars.
Unpickled void scalars must be both aligned and writeable.
BUG: Unpickled void scalars should be contiguous
|
thanks! |
Void scalars are both C- and Fortran-contiguous, so pickling and
unpickling them should result in a new void scalar that also has
these flags set. Fixes #7140.