-
-
Notifications
You must be signed in to change notification settings - Fork 5.1k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
MAINT: linalg: get rid of calc_lwork.f #3871
Conversation
Call LAPACK routines with lwork=-1 to get recommended work array sizes, instead of using calc_lwork.f (which corresponds to some old LAPACK version and may not be optimal for newer ones). Also remove the LWORK check from zgesdd wrapper --- LAPACK makes the necessary checks itself.
This is off-topic but are PR labels still needed with github's updated interface? |
The PR tag is not needed now that the interface is a bit more sane. |
MAINT: linalg: get rid of calc_lwork.f
Good riddance to |
@argriffing has already fixed |
Ripple effects: bokeh/bokeh#1695 |
It would be very good to have some early warning of things like this. Maybe some extension of the testing here:
There's some discussion about this going on at statsmodels/statsmodels#2184 |
Possibly --- also pyamg and krypy appear to use it. Can you open a |
Yes, I think we want an automated whole-stack test suite that can be easily run at least before releases. The test suite doesn't need to have the latest versions of the downstream libraries (in fact it could test several old versions too). |
The scipy-stack system does this already, but you need to do more or less manually. How about adding that to the RC process? |
Sure. Are there instructions somewhere how it works? |
I'll add those on the README, and invite you as Admin to the repo. Ralf is already an Admin. |
Filled out README.rst at https://github.com/MacPython/scipy-stack-osx-testing ; does that help? |
Ok, thanks. Quite a nice bit of infra you have there :) |
Would it make sense to somehow temporarily roll back the scipy version on pypi so it's not breaking 8,000 people's work per day? |
Surely a very fast point release would be a better option? |
To avoid this crap in the future, we could just break everything at |
In that case would it make sense to release a very fast band-aid point release, rather than a very fast point release that attempts to make the correct fix that @pv describes as a bit tricky? Maybe just add |
I think calc_lwork is used in real code in the other projects, however. |
How about applying that patch on top of 0.15.0 and releasing today as 0.15.1? |
calc_lwork in statsmodels is in an old unused function that I had forgotten about, but unfortunately the module is always imported, and breaks any standard use of statsmodels (api.py) |
I'd suggest to (i) 0.15.1 with the old+buggy calc_lwork, and (ii) get it out during the weekend and check see if something else is also on fire (I'm not going to get it done today even as a source-only release). |
We can always release 0.15.2 if anything else comes up. I can do the OSX binary builds of course, is there anything else I can help with? |
Or maybe you fixed that already for 0.14.1 and 0.15.0? Then that issue can be closed. |
+1 for underscoring everything at once and adding warnings for the regular imports. |
I added statsmodels to the scipy-stack-osx-testing test set. At the moment tests are failing because of this issue: https://travis-ci.org/MacPython/scipy-stack-osx-testing/builds/47383316 |
@matthew-brett thanks! |
Call LAPACK routines with lwork=-1 to get recommended work array sizes,
instead of using calc_lwork.f (which corresponds to some old LAPACK
version and may not be optimal for newer ones).
Also remove the LWORK check from zgesdd wrapper --- LAPACK makes the
necessary checks itself.
(This is more of a cleanup rather than a bug fix.)