casting to `quaddtype` emits uneccesary warnings

[mattip]

When using np.astype and the array has inf or nan, I see a RuntimeWarning: invalid value encountered in cast. On internal numpy float dtypes these are not emitted.

Run with python -Walways

import numpy as np
import numpy_quaddtype
a = np.arange(30, dtype=np.float32); 
a[-1] = np.inf
q = numpy_quaddtype.QuadPrecision
d = np.float64
warns = a.astype(q)
# <stdin>:1: RuntimeWarning: invalid value encountered in cast
nowarns = a.astype(d)
# This does not warn
warns.astype(d)

I think this is coming from sleef?

[SwayamInSync]

Yes it seems to be coming from sleef and also comes in during extreme cases comparison using sleef specific operations. Not 100% sure but its possible sleef using quiet comparison routines.
@seberg suggested this flag -ftrapping-math in this comment numpy/numpy-user-dtypes#121 (comment) (It is not yet applied in the sleef's meson build) I'll look into it

[juntyr]

@SwayamInSync Is this still an issue?

[SwayamInSync]

Yes this still persist and while working on nextafter I kind of think sleef does not sets FPE at all, as if you remember the checks I kept at last (in the code for nextafter) there originally I was handling them but removed later when doing PR.

I will confirm my suspicsion after all ufunc gets completed and see if the setting up flag would be enough or we have to do this manually

[Brian-Castle]

I have a question about this. If I cast a quaddtype that's out of 64 bit range, am I going to get a nan there? How about, a number that's so small it results in 0 in 64-bit land? Does an underflow result in Nan or just 0? Is the behavior consistent between Python and C?

[SwayamInSync]

quaddtype -> flaot64 (out of range) => +-inf
quad too small for float64 => Just 0, never NaN. Underflow is sign-preserving

And yeah this behaviour is consistent between Python and C. (both follow IEEE-754)

Just for integer casting there were inconsistencies we fixed that in #104 which isn't on pypi releases yet.

[Brian-Castle]

Thank you! Will this work with PyTorch tensors? dtype=torch.QuadDtype?

[ngoldbaum]

Thank you! Will this work with PyTorch tensors? dtype=torch.QuadDtype?

No, this is a NumPy dtype and has nothing to do with PyTorch. PyTorch doesn't automatically interoperate with all NumPy dtypes like that.

I'm also not sure how one would even use quad floats on a GPU, since GPUs don't have hardware quad support.

[Brian-Castle]

Okay. How about the structured arrays in numpy, can we use 'fQuad' or np.QuadDtype? Like this:

np.dtype([('x', 'f4'), ('y', np.float32), ('z', 'f4', (2, 2))])

[SwayamInSync]

Structured arrays work but you have to pass the QuadPrecDType() instance instead of string identifiers.

[Brian-Castle]

Thank you!