Add dtype and manipulation functions #1566
Conversation
|
View rendered docs @ https://intelpython.github.io/dpnp/pull/1566/index.html |
|
It seems we do not to skip the fallowing tests anymore. tests/third_party/cupy/manipulation_tests/test_dims.py::TestBroadcast_param_0_{shapes=[(), ()]}::test_broadcast_arrays |
|
@npolina4 |
|
The below example works differently with numpy and description: a = numpy.ones((2, 3), dtype='i4')
b = a.astype('i4', copy=False)
b is a
# Out: True
a = dpnp.ones((2, 3), dtype='i4')
b = a.astype('i4', copy=False)
b is a
# Out: False |
| Parameters | ||
| ---------- | ||
| dtype : dtype | ||
| Target data type. |
There was a problem hiding this comment.
What is about None value? It would be good to state clearly, since dpctl and numpy supports that but might behave differently.
| return dpt.astype(x1, dtype, order=order, casting=casting, copy=copy) | ||
| Parameters | ||
| ---------- | ||
| dtype : dtype |
There was a problem hiding this comment.
Need to describe x1 and casting
| Otherwise, a copy is returned. | ||
| Parameters | ||
| ---------- | ||
| dtype : dtype |
There was a problem hiding this comment.
Need to describe casting
| Limitations | ||
| ----------- | ||
| Parameter `subok` is supported with default value. | ||
| Otherwise the function will be executed sequentially on CPU. |
There was a problem hiding this comment.
Probably it would be better to get rid of the falling back on CPU and to raise an exception explicitly, like cupy does?
| new_array = self.__new__(dpnp_array) | ||
| new_array._array_obj = dpt.astype( | ||
| self._array_obj, dtype, order=order, casting=casting, copy=copy | ||
| return dpnp.astype( |
There was a problem hiding this comment.
It would be better to have main implementation here and then dpnp.astype will be equal to
if dpnp.is_supported_array_type(x1):
return x1.astype(...)
else:
raise TypeError(...)
Also it will allow us to avoid invoking dpnp.get_usm_ndarray(...) and dpnp_array._create_from_usm_ndarray(...).
| @@ -150,30 +151,41 @@ def asnumpy(input, order="C"): | |||
|
|
|||
|
|
|||
| def astype(x1, dtype, order="K", casting="unsafe", subok=True, copy=True): | |||
There was a problem hiding this comment.
Neither numpy or cupy has astype method (they have it only in the context of ndarray), so I would propose to remove unsupported subok parameter here.
| ------- | ||
| out : dpnp.array | ||
| If ``copy`` is False and no cast is required, then the array itself is returned. | ||
| Otherwise, it returns a (possibly casted) copy of the array. |
There was a problem hiding this comment.
It sounds a bit unclear what "possibly casted" means. Could you please clarify that in the description or rephrase it.
In general, numpy and dpctl returns a view where possible.
| @@ -597,32 +597,46 @@ def asnumpy(self): | |||
| return dpt.asnumpy(self._array_obj) | |||
|
|
|||
| def astype(self, dtype, order="K", casting="unsafe", subok=True, copy=True): | |||
| """Copy the array with data type casting. | |||
| """ | |||
There was a problem hiding this comment.
Could you please check tests/third_party/cupy/core_tests/test_ndarray_copy_and_view.py::TestArrayCopyAndView and unmute where possible?
Also it would be great refresh test_ndarray_copy_and_view.py file.
RP will separate:
#1582 Leverage dpctl.tensor.iinfo() and dpctl.tensor.finfo() implementation.
#1583 Clean up old implementation of logic functions
#1585 Updated atleast_2d and atleast_3d functions
#1586 Implemented dpnp.tile function.
Add dtype and manipulation functions and tests.
Manipulation functions: dpnp.broadcast_arrays, dpnp.repeat, dpnp.vstack, dpnp.ravel
Data Type functions: dpnp.can_cast