BUG: make result of isfinite/isinf/signbit a boolean #6994
Conversation
|
let me mess a bit with this |
|
Yeah, it's painful. I have another test running to see if anything changed between merging your branch and current master. |
|
lets see if I can reproduces it in my windows vm, probably faster |
|
timeouts oO |
|
Well, I was able to restart all of them... |
|
hm it works on win32 python34, so maybe a python3.5 problem? Unfortunately I can't test that as I only have windows xp available right now on which 3.5 won't install ._. |
|
Curiously, when I revert back to merge, only |
|
Looks like you didn't revert the whole merge, it's now failing on 64 bits also.. I did |
|
yes that was intentional, so its not the vector code that narrows it down |
|
Can't rule out compiler bugs either. I think it safe to say that this problem is Python 3.5 specific, which includes new compilers and possibly build flags. |
|
The half-float error vaguely smells of a possible optimization problem. |
|
ah I think I know the issue! |
|
lets see if that works, been using too much c99/c++ lately to notice this error ... |
may return something else than one or zero and npy_bool is unfortunately an int8 not a c99 bool
|
If only we could make c99 standard... Although I expect that would turn up other problems. |
|
So, this one checked out ( https://ci.appveyor.com/project/charris/numpy/build/1.0.446 ). Did it just resolve down to adding |
|
@jakirkham Looks like it. |
BUG: make result of isfinite/isinf/signbit a boolean
|
@juliantaylor Kudos for tracking that down. |
|
Fixes #6991 |
This reverts commit 4b82b1b, reversing changes made to 3a9c90c. First of two reversions to undo vectorization of isinf, isnan, and signbit for Numpy 1.11.0. The changes led to test failures on windows for Python 2.6. Because Python 2.6 will not be supported by Numpy 1.12, this does not need to be done for current master.
may return something else than one or zero and npy_bool is
unfortunately an int8 not a c99 bool