Fedora build patches #9304
Fedora build patches #9304
Conversation
These are either deprecations, or checks for old, but probably incorrect, behaviour.
For some reason, NaN gets converted to 0 as an integer instead of INT_MIN like on x86. Fixes matplotlib#6538.
|
@QuLogic are you planning to add more to this? |
|
I was really hoping to figure out the ppc64 barb issue over the weekend, but as I'm not at a suitable computer at the moment, and have other things to do, I guess that won't be possible. I believe the remaining build issues are simply due to FreeType 2.7+ and small precision things which we can probably ignore for the time being. Feel free to merge this whenever it's approved. |
| @@ -603,7 +603,8 @@ def test_load_from_url(): | |||
|
|
|||
| @image_comparison(baseline_images=['log_scale_image'], | |||
| remove_text=True) | |||
| def test_log_scale_image(): | |||
| # The recwarn fixture captures a warning in image_comparison. | |||
| def test_log_scale_image(recwarn): | |||
There was a problem hiding this comment.
I think this is throwing a warning that we should try and fix instead of ignoring it. Presumably ax.imshow sets axis=equal behind the scenes, which causes ax.set_yscale('log') to throw a warning. If this is all valid code (as I think it is) I don't think it should throw a warning.
There was a problem hiding this comment.
But how could we catch it? Setting equal aspect ratio and setting log scaling are independent, and I think there's no link saying that equal aspect ratio was requested by imshow.
There was a problem hiding this comment.
In principle we could set the aspect to 'auto' after calling imshow but I would expect that to change the output. I am 👍 on just catching this warning.
|
So looking at the C standard 6.3.1.4 paragraph 1:
I'm not sure if NumPy just defers to C casts there, but I guess if it does, that sort of explains why the cast in |
|
Probably worth raising this with numpy? Obviously unlikely to help now, but I think it would make sense for the numpy error system to report such overflows/UB... |
|
There appear to be a few upstream reports about this; numpy/numpy#8325 seems farthest along and they seem to have come to the same root cause. |
|
Wait, this was not merged into master? |
PR Summary
Patches needed to build & test on Fedora on multiple arches, with the exception of the deprecation warning capture (which is just nice to have because of the pytest bug). Of course, this does not include the tolerance bump required to build against FreeType 2.7.1 or 2.8.
This works on x86_64 and arm. I need to go through the results of aarch64/ppc64/ppc64le as well.
PR Checklist