Add Cupy space

[adler-j]

This should replace #1231. I'm currently in the middle of a nightmarish merge of master changes and changes in this branch.

Edit: Overall my reaction to this PR atm is: Oh god please kill me. But I'll try to persevere.

[pep8speaks]

Checking new PR...

• In the file odl/space/cupy_tensors.py, following are the PEP8 issues :

Line 1869:21: E123 closing bracket does not match indentation of opening bracket's line

• Complete extra results for this file :

                ):                    ^---

• In the file odl/util/testutils.py, following are the PEP8 issues :

Line 30:1: E302 expected 2 blank lines, found 1

• Complete extra results for this file :

def _ndigits(a, b, default=None):^---

• In the file odl/util/utility.py, following are the PEP8 issues :

Line 52:5: E123 closing bracket does not match indentation of opening bracket's line

• Complete extra results for this file :

)    ^---

[kohr-h]

Is it mostly the tests or other stuff as well?

[kohr-h]

Just a minor remark: We should require CuPy >= 4.0 since it's the only version available through conda (I heard they also work on Windows support for 5.0), and since it supports a lot of useful stuff that earlier versions don't, most of all FFTs.

[adler-j]

Basically someone (you and me) has done the same work exactly twice, one on master and one here. So there are merge conflicts everywhere. Join that with this branch actually containing some fundamental changes and you have a nightmare merge. I predict it will take at least a few days before it's all running smoothingly.

W.r.t windows its already up, I'm using it atm.

[kohr-h]

Basically someone (you and me) has done the same work exactly twice, one on master and one here. So there are merge conflicts everywhere. Join that with this branch actually containing some fundamental changes and you have a nightmare merge. I predict it will take at least a few days before it's all running smoothingly.

😬 😇

W.r.t windows its already up, I'm using it atm.

Cool, I guess through pip?

[adler-j]

Question for @kohr-h. At what point (if at all) did np.absolute(complex_vector) start returning something in complex_vector.space.real_space?

[kohr-h]

This can only be the effect of __array_ufunc__.

[adler-j]

Is there a list of the ufuncs that do this somewhere?

[kohr-h]

>>> ufuncs = list(
...     filter(
...         lambda a: isinstance(getattr(np, a), np.ufunc), 
...         dir(np)
...     )
... )
>>> c2r = list(
...     filter(
...         lambda f: any(t in getattr(np, f).types 
...                       for t in('F->f', 'D->d', 'G->g')),
...         ufuncs,
...     )
... )
>>> c2r
['abs', 'absolute']

[adler-j]

Thanks!