dpctl could not properly handle NaN edges in `clip` function

[antonwolfy]

dpctl raises an exception for either any edge as NaN or both edges as NaN passed to dpctl.tensor.clip function.
While NumPy accept any of below combinations without an issue:

a = numpy.arange(7)

numpy.clip(a, numpy.nan, None)
numpy.clip(a, None, numpy.nan)
numpy.clip(a, numpy.nan, numpy.nan)
numpy.clip(a, -2, numpy.nan)
numpy.clip(a, numpy.nan, 10)

The below is example of errors raised by dpctl:

dpctl.__version__
# Out: '0.15.1dev2+28.gebf94b04cf'

# create an array
a = dpt.arange(7)

# clip with any edge as NaN, where another is None
dpt.clip(a, numpy.nan, None)
---------------------------------------------------------------------------
TypeError Traceback (most recent call last)
Cell In[3], line 1
----> 1 dpt.clip(a, numpy.nan, None)

File ~/miniconda3/envs/dpnp_dev/lib/python3.9/site-packages/dpctl/tensor/_clip.py:433, in clip(x, min, max, out, order)
 429 raise ValueError(
 430 "only one of `min` and `max` is permitted to be `None`"
 431 )
 432 elif max is None:
--> 433 return _clip_none(x, min, out, order, tei._maximum)
 434 elif min is None:
 435 return _clip_none(x, max, out, order, tei._minimum)

File ~/miniconda3/envs/dpnp_dev/lib/python3.9/site-packages/dpctl/tensor/_clip.py:249, in _clip_none(x, val, out, order, _binary_fn)
 247 _fp16 = sycl_dev.has_aspect_fp16
 248 _fp64 = sycl_dev.has_aspect_fp64
--> 249 if not _can_cast(val_dtype, res_dt, _fp16, _fp64):
 250 raise ValueError(
 251 f"function 'clip' does not support input types "
 252 f"({x_dtype}, {val_dtype}), "
 253 "and the inputs could not be safely coerced to any "
 254 "supported types according to the casting rule ''safe''."
 255 )
 257 orig_out = out

File ~/miniconda3/envs/dpnp_dev/lib/python3.9/site-packages/dpctl/tensor/_type_utils.py:103, in _can_cast(from_, to_, _fp16, _fp64)
 98 def _can_cast(from_: dpt.dtype, to_: dpt.dtype, _fp16: bool, _fp64: bool):
 99 """
 100 Can `from_` be cast to `to_` safely on a device with
 101 fp16 and fp64 aspects as given?
 102 """
--> 103 can_cast_v = dpt.can_cast(from_, to_) # ask NumPy
 104 if _fp16 and _fp64:
 105 return can_cast_v

File ~/miniconda3/envs/dpnp_dev/lib/python3.9/site-packages/dpctl/tensor/_manipulation_functions.py:706, in can_cast(from_, to, casting)
 701 raise TypeError("Expected dtype type.")
 703 dtype_to = dpt.dtype(to)
 705 dtype_from = (
--> 706 from_.dtype if isinstance(from_, dpt.usm_ndarray) else dpt.dtype(from_)
 707 )
 709 _supported_dtype([dtype_from, dtype_to])
 711 return np.can_cast(dtype_from, dtype_to, casting)

TypeError: Tuple must have size 2, but has size 1

# clip with both edges as NaNs
dpt.clip(a, numpy.nan, numpy.nan)
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[10], line 1
----> 1 dpt.clip(a, numpy.nan, numpy.nan)

File ~/miniconda3/envs/dpnp_dev/lib/python3.9/site-packages/dpctl/tensor/_clip.py:531, in clip(x, min, max, out, order)
 523 buf1_dt, buf2_dt, res_dt = _check_clip_dtypes(
 524 x_dtype,
 525 min_dtype,
 526 max_dtype,
 527 sycl_dev,
 528 )
 530 if res_dt is None:
--> 531 raise ValueError(
 532 f"function '{clip}' does not support input types "
 533 f"({x_dtype}, {min_dtype}, {max_dtype}), "
 534 "and the inputs could not be safely coerced to any "
 535 "supported types according to the casting rule ''safe''."
 536 )
 538 orig_out = out
 539 if out is not None:

ValueError: function '<function clip at 0x7fade5dad310>' does not support input types (int64, float32, float32), and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''.

# clip with any edge as NaN, where another is a number
dpt.clip(a, -2, numpy.nan)
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
Cell In[12], line 1
----> 1 dpt.clip(a, -2, numpy.nan)

File ~/miniconda3/envs/dpnp_dev/lib/python3.9/site-packages/dpctl/tensor/_clip.py:531, in clip(x, min, max, out, order)
 523 buf1_dt, buf2_dt, res_dt = _check_clip_dtypes(
 524 x_dtype,
 525 min_dtype,
 526 max_dtype,
 527 sycl_dev,
 528 )
 530 if res_dt is None:
--> 531 raise ValueError(
 532 f"function '{clip}' does not support input types "
 533 f"({x_dtype}, {min_dtype}, {max_dtype}), "
 534 "and the inputs could not be safely coerced to any "
 535 "supported types according to the casting rule ''safe''."
 536 )
 538 orig_out = out
 539 if out is not None:

ValueError: function '<function clip at 0x7fade5dad310>' does not support input types (int64, int64, float32), and the inputs could not be safely coerced to any supported types according to the casting rule ''safe''.

[ndgrigorian]

Some of this behavior is expected with the current implementation of clip: it does not perform type promotion (where Numpy does), and NaNs cannot be cast to exact data types. inf and -inf will have similar issues.

The array API however seems to have come to the conclusion that type promotion is the correct approach to clip among other changes. Those changes will remove this limitation.

The case with the unhelpful TypeError: Tuple must have size 2, but has size 1 error will also need to be examined.