Improve CI

[drahnr]

Currently execution is only done on native, but this shall not limit test compilation for cudnn.

cargo b --tests

for all feature sets

should suffice to catch most if not all issues with compilation.

Runtime errors are hard to catch, since native execution with running GPUs is hard to achieve with concourse, and there is nothing else out there that one can host yourself with the same feature set.

[drahnr]

Related to #24

[drahnr]

Also note the issue that cargo does not handle feature flags with workspaces just yet, which makes that rather tedious to implement in CI right now.

[drahnr]

Also note the issue that cargo does not handle feature flags with workspaces just yet, which makes that rather tedious to implement in CI right now.

The related cargo issue: rust-lang/cargo#5015

[drahnr]

Container runtime with Nvidia
https://devblogs.nvidia.com/gpu-containers-runtime/
containers/podman#884

[drahnr]

From: https://concourse-ci.org/concourse-worker.html#configuring-gdn-server

/usr/local/concourse/bin/gdn server --help

shows --runtime-plugin= which defaults to runc :) so plugging in nvidia-container-runtime should do the trick 🤞
If not, setting a prestart hook on the default shipped runc should work too via --runtime-plugin-extra-arg= 🤞

[drahnr]

CI is now working, the setup complexity got a little more intricate than anticipated, but it works now (and showed that we might need to regenerated the cudnn-sys with bindgen, see https://ci.spearow.io/teams/spearow/pipelines/juice/jobs/pr-test-juice/builds/114)