Array Casting Support for fixed-length String dtypes

[SwayamInSync]

import numpy as np
from numpy_quaddtype import *

a = np.array([QuadPrecision("1.23456789012345678901234567890123456")])
a.astype(str)  # Raises TypeError

Traceback (most recent call last):
  File "main.py", line 39, in <module>
    a.astype(str)  # Raises TypeError
    ^^^^^^^^^^^^^
TypeError: cannot cast dtype QuadPrecDType(backend='sleef') to <class 'numpy.dtypes.StrDType'>.

[SwayamInSync]

UPDATE:
unsupported types: [np.str_, np.bytes_, str, bytes]

[SwayamInSync]

I didn't inspect yet, but implementing this might be non-traditional as the string can be of any size + whether the numpy keeps the track of correct stride.

@ngoldbaum maybe you have any pointers regarding this

[ngoldbaum]

With the fixed-width strings you'll just truncate to the width of the string if the output string doesn't fit in the width of the string.

For StringDType I can point you at lots of examples. You could look at the StringDType cast implementations to and from double for inspriation. I think we just cast long double to double, so it'd be nice to point here if anyone cares about that.

[SwayamInSync]

I see, so to ensure round trip 36 is usually enough (also assuming scientific notation) but internally we keep the buffer of 128 size so just for consistency will go with this.

You could look at the StringDType cast implementations to and from double for inspriation

They are present in the user-dtype repo or inside the NumPy?

[ngoldbaum]

inside numpy:

https://github.com/numpy/numpy/blob/41c06bc51f5848925ad647bd8dfd1bfc7aee4f64/numpy/_core/src/multiarray/stringdtype/casts.cpp#L1092

https://github.com/numpy/numpy/blob/main/numpy/_core/src/multiarray/stringdtype/casts.cpp#L1198

and now that I look, it seems longdouble is handled correctly now. I must have been misremembering the old state before that was fixed.

The version of stringdtype in this repo is the prototype and is seriously out of date.

[SwayamInSync]

Ah, StringDType() does not inherit np.str_ so creating QuadPrecision scalar currently has 0 support for it.
Although np.str_ and others work.

In [3]: print(StringDType.__mro__)
(<class 'numpy.dtypes.StringDType'>, <class 'numpy.dtype'>, <class 'object'>)

In [4]: print(np.str_.__mro__)
(<class 'numpy.str_'>, <class 'str'>, <class 'numpy.character'>, <class 'numpy.flexible'>, <class 'numpy.generic'>, <class 'object'>)

I maybe open a new issue for complete support for StringDType