`dpctl.tensor.result_type` raises `ValueError`

[vtavana]

The following example raises a ValueError in dpctl while it works fine in NumPy.

import dpctl.tensor as dpt, numpy
dtype = numpy.dtype(numpy.float64)
scalar = dtype.type(3)
numpy.result_type(scalar, scalar)
# dtype('float64')

>>> dpt.result_type(scalar, scalar)    
# File ~/micromamba/envs/dpnp_2024/lib/python3.11/site-packages/dpctl/tensor/_type_utils.py:785, in result_type(*arrays_and_dtypes)
#    783         res_dt = _to_device_supported_dtype_impl(res_dt, has_fp16, has_fp64)
#    784 else:
#--> 785     res_dt = np.result_type(*dtypes)
#    786     if weak_dtypes:
#    787         weak_dt_obj = [wdt.get() for wdt in weak_dtypes]
#ValueError: at least one array or dtype is required

[oleksandr-pavlyk]

Don't you think this is as designed? dpctl can not combine two scalars, and so it is reasonable for result_type to raise for such inputs as well.

[vtavana]

Don't you think this is as designed? dpctl can not combine two scalars, and so it is reasonable for result_type to raise for such inputs as well.

It does not raise an error for all the cases, for instance, it works fine with float32.

import dpctl.tensor as dpt, numpy
dtype = numpy.dtype(numpy.float32)
scalar = dtype.type(3)
dpt.result_type(scalar, scalar) 
# dtype('float32')

[oleksandr-pavlyk]

I think the real issue is that dpt.result_type does not raise if none of the arguments is an array. I intend to fix that.

The discrepancy you found with np.float32 and np.float64 comes from np.dtype(np.float64).type being Python native built-in type float, whereas np.dtype(np.float32).type is a custom scalar type np.float32.

dpt.result_type treats Python scalars bool, int, float and complex as weak types as per NEP-0050, but treats other types as possible inputs to dpt.dtype constructor.

[oleksandr-pavlyk]

Ok, the result_type implementation has an issue when all inputs of weak dtypes, i.e. Python scalars

In [2]: dpt.result_type(1, 1.0)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[2], line 1
----> 1 dpt.result_type(1, 1.0)

File ~/repos/dpctl/dpctl/tensor/_type_utils.py:783, in result_type(*arrays_and_dtypes)
    781         res_dt = _to_device_supported_dtype_impl(res_dt, has_fp16, has_fp64)
    782 else:
--> 783     res_dt = np.result_type(*dtypes)
    784     if weak_dtypes:
    785         weak_dt_obj = [wdt.get() for wdt in weak_dtypes]

ValueError: at least one array or dtype is required