Disallow numeric scalar conversation for non-0D arrays #2223
Conversation
|
View rendered docs @ https://intelpython.github.io/dpctl/pulls/2223/index.html |
|
Array API standard conformance tests for dpctl=0.22.0dev0=py310h93fe807_103 ran successfully. |
|
Array API standard conformance tests for dpctl=0.22.0dev0=py310h93fe807_108 ran successfully. |
| cdef inline void _check_0d_scalar_conversion(object usm_ary) except *: | ||
| "Raise TypeError if array cannot be converted to a Python scalar" | ||
| if (usm_ary.ndim != 0): | ||
| raise TypeError( |
There was a problem hiding this comment.
| raise TypeError( | |
| raise ValueError( |
There was a problem hiding this comment.
I guess NumPy raises a TypeError but this really seems like a ValueError to me
| assert not X.__le__(-1) | ||
| assert not X.__eq__(-1) | ||
| assert X.__ne__(-1) | ||
| assert dpt.all(X.__gt__(-1)) |
There was a problem hiding this comment.
instead of adding these additional kernel launches, we can just either strip the dimension off of X or we can use dpt.asarray above, so that X is an array scalar
| x2 = dpt.ones([1], dtype="i1") | ||
| r2 = dpt.argsort(x2, kind="radixsort") | ||
| assert r2 == 0 | ||
| assert dpt.all(r2 == 0) |
There was a problem hiding this comment.
similar to the other test, we can save a kernel launch by stripping the dimension off of r2. It's not too bad of a hack I feel.
ndgrigorian
left a comment
ndgrigorian
left a comment
There was a problem hiding this comment.
aside from minor comments this LGTM
4 participants
This PR partially aligns
usm_ndarrayscalar conversion semantics with the Python Array API by allowing numeric scalar conversion only for 0D usm_ndarraysConversions to Python numeric scalars via
int,float,complexand their corresponding methods (__int__,_float__,__complex__) now raiseTypeErrorwhen applied to usm_ndarrays withndim !=0Note:
For full compliance with the Python Array API, the same check should be used for truth values
__bool__method and integer indexing index methodBut these changes would impact dpnp compatibility with Numpy 2.4 which currently disallows only numeric scalar conversation (a minor behavioral difference I think)
In addition several
dpctl testsrely on implicit truth conversion for non-0D arrays and must be updated accordingly