Casting to integer on conversion

[znicholls]

Hi there,

I have a question about behaviour which I was hoping you could help with. It focuses on this comment:

cf-units/cf_units/__init__.py

Line 1845 in a3eeaf1

# Preserve the datatype of the input array if it was float32.

My question is: why isn't this also applied to np.float64 datatypes?

More info:

The minimal example below shows how this can cause a np.float64 dtype to be cast to integer even though a np.float32 dtype is not. This seems surprising, but maybe that is the idea? (As a sidenote, this only appears to happen on a linux operating system, it doesn't happen on mac...)

In [1]: import cf_units

In [2]: import numpy as np

In [3]: cf_units.__version__
Out[3]: '2.1.4'

In [4]: start = cf_units.Unit('days since 2100-01-01', calendar='365_day')

In [5]: target = cf_units.Unit('days since 2006-01-01', calendar='365_day')

In [6]: start.convert(np.array([1, 2, 3]).astype(np.float32), target)
Out[6]: array([34311., 34312., 34313.], dtype=float32)

In [7]: start.convert(np.array([1, 2, 3]).astype(np.float32), target).dtype
Out[7]: dtype('float32')  # type preserved

In [8]: start.convert(np.array([1, 2, 3]).astype(np.float64), target)
Out[8]: array([34311, 34312, 34313])

In [9]: start.convert(np.array([1, 2, 3]).astype(np.float64), target).dtype
Out[9]: dtype('int64')  # type not preserved

Thanks!

[znicholls]

Doing more digging, this appears to be a result of Unidata/cftime#225

[znicholls]

@ocefpaf are you the right person to ask about this? If no, is it possible to point me in the direction of the person who is please?