search

exoplanet-dev / exoplanet

Fast & scalable MCMC for all your exoplanet needs!
https://docs.exoplanet.codes
MIT License
206 stars 52 forks source link

CI builds on Windows fail intermitently #92

Closed dfm closed 4 years ago

dfm commented 4 years ago

The Windows builds on GitHub actions fail (seemingly) randomly with the following assembly error:

Error: invalid register for .seh_savexmm

If I re-run the build sometimes exactly the same code will compile just fine. I've tried all of the suggestions from this stack overflow question, but none of them work:

I think this must have something to do with how Theano configures the compile flags, but I haven't been able to track it down.

If anyone has access to a Windows machine and/or some experience with such things, we'd love some help!!

jobovy commented 4 years ago

Not directly helpful, but I've been using VirtualBox to develop and maintain the Windows version of galpy and that's been very helpful in iterating on lots of compilation issues. If I recall, it takes a little effort to set it up for development (but installing conda is pretty easy), but once it's set up, you can quickly boot it up to diagnose issues.

Also not directly helpful, but I've found using Miniconda on GH Actions to often help provide a more stable testing environment.

dfm commented 4 years ago

@jobovy: Thanks for the tips!

I'll try out VirtualBox - I have a few projects where I would like to make sure that the the Windows version is stable so this will come in very handy!

I was originally using miniconda, but tried switching to the bundled version when I was getting this same issue there.

dfm commented 4 years ago

Quick update: Everything seems to run and compile just fine on VirtualBox :/ and I confirmed that the issues still remain when I naively use miniconda.

On Twitter, @ahartikainen suggested:

Try to use conda-forge only env

conda create -n myenv python=3.7 -c conda-forge and then activate and install only from conda-forge (+pip)

I wasn't able to solve a consistent environment using only conda-forge (a few lower-level packages seem to be required from defaults) but I'm giving it a shot and I'll report back.

dfm commented 4 years ago

It looks like the conda-forge suggestion might have worked (#93) so I'm going to close this for now and re-open if the failures start again. Thanks all for your suggestions!

dfm commented 4 years ago

Re-opening because this fix didn't work!

ahartikainen commented 4 years ago

Is this related

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=65782

https://github.com/numpy/numpy/issues/14787

Would old numpy work?

dfm commented 4 years ago

I came across those issues as well and it seems like they must be related, but the recommended fixes didn't work for me in general. I'll play around with numpy versions to see if that helps. Thanks!