np.place segfaults for quaddtype

[juntyr]

import numpy as np
from numpy_quaddtype import QuadPrecDType

a = np.zeros((18, 1), dtype=QuadPrecDType())
m = np.zeros((18, 1), dtype=np.bool)
m[3:-3] = True
v = np.ones(12, dtype=QuadPrecDType())

np.place(a, m, v)

print(a, m, v)

never reaches the print in numpy-quaddtype==0.2.0.

@SwayamInSync has this been fixed since?

[SwayamInSync]

This requires the implementation of copyswap otherwise if urgent then a similar operation np.copyto(arr, vals, where=mask) works

[SwayamInSync]

@seberg just a heads up on another function that relies on copyswap
and inside dtypemeta.c it assigns NULL to that function pointer leading to segfault

[juntyr]

This requires the implementation of copyswap otherwise if urgent then a similar operation np.copyto(arr, vals, where=mask) works

np.copyto is unfortunately not the right workaround since np.place takes a flat differently sized vals array. However, np.put(arr, np.flatnonzero(mask), vals) seems to work

[juntyr]

@SwayamInSync This should definitely be fixed, but since there is a workaround it's not urgent (can be included in the upcoming big release)

[seberg]

Yeah, should at least raise a nice error. For most use-cases, I think a[m] = v is nicer (yes, place has some funny semantics).

[juntyr]

@SwayamInSync This is still not fixed in the current main branch version

[SwayamInSync]

Yes this and other issues makred at #bugs are depending on the same changes to be made on numpy + adding support to different endian casting within quaddtype

So this might take a while and not that critical as a workaround exists, we will cover this in future releases

[juntyr]

Ok, thanks for the update! I pulled some wheels from CI to test v1.0 with my downstream project and removed all bug-workarounds from earlier; this is the only one I still need but everything else seems to be working so far