BUG (I guess?): Make a @= b error out #6000
Conversation
|
Doh, right, there's actually no way to test this at all except on py3.5. Our current |
43d7e90 to
3859b2f
Compare
|
Just pushed a new version of this that's ready for review. I tested 3859b2f locally against a recent python 3.5 build, and the new test passed. |
| # we avoid writing the token `exec` so as not to crash python 2's | ||
| # parser | ||
| exec_ = getattr(builtins, "exec") | ||
| assert_raises(TypeError, exec_, "a @= b", globals(), locals()) |
There was a problem hiding this comment.
Could also use operator.imatmul probably, though I am not sure I care.
There was a problem hiding this comment.
Eh, sure, I'll throw that in too :-)
There was a problem hiding this comment.
I thought to avoid the exec dance ;p, but nvm. All is good in any case.
|
Anyway, seems the right thing to do to just not support it for the moment. For stacked matrices, it would probably be easy to support (and make sense to safe memory), since einsum supports it, and I think we currently use einsum (for non-stacked, a simple assignment using |
|
Nvm. it is not possible to support even with einsum, since it should be (at least almost) impossible to force correct/large enough buffering. (It might be that forcing buffering + iteration size would do, but I don't even think einsum exposes that). |
Before this change, we defined a nb_matrix_multiply slot but not a nb_inplace_matrix_multiply slot, which means that a statement like a @= b would be silently expanded by the CPython interpreter to become a = a @ b This is undesireable, because it produces unexpected memory allocations, breaks view relationships, and so forth. This commit adds a nb_inplace_matrix_multiply slot which simply errors out, and suggests that users write 'a = a @ b' explicitly if that's what they want.
3859b2f to
1adcdf7
Compare
|
Addressed all comments. |
|
Thanks. |
Before this change, we defined a
nb_matrix_multiplyslot but not anb_inplace_matrix_multiplyslot, which means that a statement likewould be silently expanded by the CPython interpreter to become
This is undesireable, because it produces unexpected memory
allocations, breaks view relationships, and so forth.
This commit adds a
nb_inplace_matrix_multiplyslot which simply errorsout, and suggests that users write
a = a @ bexplicitly if that'swhat they want.
[Noticed this while working on a
__numpy_ufunc__patch, so here's a quick fix that I haven't even tested yet :-). Let's see what Travis says...]