Feature: bitcast to an equal-sized unsigned integer type

[juntyr]

Bitcasting (via numpy.view) from floating point to integer is incredibly useful, but currently not possible since there is no 128bit integer type in numpy.

[SwayamInSync]

You can use structured dtype (2 int64) in numpy to view quaddtype (similar to what we do in the codebase)

In [9]: x = QuadPrecision(2.5)

In [10]: dt = np.dtype([('low', np.uint64), ('high', np.uint64)])

In [11]: x.view(dt)
Out[11]: np.void((0, 4611756387171565568), dtype=[('low', '<u8'), ('high', '<u8')])

Here,

• lower bits are 0
• Higher bits are 4611756387171565568

Now you can extract the exponent and mantissa bits

>>> hex(4611756387171565568)
'0x4000400000000000'

• Sign bit: 0 -> positive
• Exponent bits: 0x4000
• Stored mantissa (high 48 bits): 0x400000000000

[SwayamInSync]

Hey @juntyr let me know if this helps and we can close this.
I am not sure what else we can possibly do here

[juntyr]

Thanks @SwayamInSync! I think I’d like to keep the issue open until (perhaps at some point) there is a uint128 dtype for numpy. While casting to uint64x2 is a good step, you then still have to implement all operations yourself, e.g. how to add one (which requires an add with carry across, which I don’t know how to do in numpy)