set CF_CUDA_ENABLED=True if compiler('cuda') is detected in meta.yaml

[h-vetinari]

From the discussion in conda-forge/conda-forge-pinning-feedstock#1708, and particularly, this description of what to do.

I couldn't think of a way to detect compiler('cuda'), because - AFAICT - the ingestion of meta.yaml is done a couple lines above by conda_build.api.render, and after that, the jinja-functions will have already been filled in.

I think it's a feasible workaround though to just use r"^nvcc.*" (as a build requirement) as a stand-in for detecting that.

I haven't yet written a test for it, this is more to check whether I'm on the right path.

CC @isuruf @jakirkham

[isuruf]

Rendering and then setting the env variable is too late. You have to set the env variable at the same time that DEFAULT_LINUX_VERSION is set. You can read meta.yaml and do a text search instead.

[h-vetinari]

Ping @isuruf, I added a test and it's green now. :)

[isuruf]

The test doesn't really test that the CF_CUDA_ENABLED is set during rerendering time. The easiest way to check is to add a skip like skip: True # [os.environ.get("CF_CUDA_ENABLED") == "True"] and see if the build was skipped.

[h-vetinari]

The test doesn't really test that the CF_CUDA_ENABLED is set during rerendering time. The easiest way to check is to add a skip like skip: True # [os.environ.get("CF_CUDA_ENABLED") == "True"] and see if the build was skipped.

I didn't understand the test design here very well (for example it seems most tests brings their own fixture?), so I tried to follow the test for DEFAULT_LINUX_VERSION. Are you saying I should add another test with another fixture (containing the skip you mentioned), and validate that there's no cuda-output?

Or - since you approved the changes - that this is not necessary now?

[h-vetinari]

Thanks for the corrections! I cannot commit them through the GH UI (not enough rights here), but if you want I can recreate them locally?

[isuruf]

Can you add a news entry?