[CI] build-and-test-differential only tests CUDA

[ambroise-arm]

Checklist

• I've read the contribution guidelines.
• I've searched other issues and no duplicate issues were found.
• I'm convinced that this is not my fault but a bug.

Description

The build-and-test-differential workflow only starts a runner with cuda enabled:

autoware.universe/.github/workflows/build-and-test-differential.yaml

Line 17 in 03de605

- -cuda

This prevents catching issues on systems that don't have cuda enabled.
For example #2370 successfully passed CI, but failed once merged: https://github.com/autowarefoundation/autoware.universe/actions/runs/3583250712/jobs/6028456850

Expected behavior

CUDA shouldn't be required to build Autoware

Actual behavior

The CI for PRs doesn't check that.

Steps to reproduce

Build current tip of main (03de605) on a system without cuda.

Versions

No response

Possible causes

No response

Additional context

No response

[ambroise-arm]

cc @kenji-miyake

[kenji-miyake]

@ambroise-arm Actually, it was intentionally removed temporarily.
#1669

After we move to Humble and remove Galactic CI workflows by #2408, I believe we can re-enable no-cuda checks.

[kenji-miyake]

Ah, sorry. I might misunderstand your problem. Will fix my comment above.

[kenji-miyake]

Okay, I think my comment wasn't so wrong.

CUDA shouldn't be required to build Autoware

I believe it's not true. We need CUDA for now because autoware.universe depends on CUDA yet.
But the no-cuda environment should also be tested.

@ambroise-arm What do you think?

[ambroise-arm]

I believe your comment gives valuable information as it is.
Is the re-enablement of the no-cuda checks tracked somewhere? If not we could keep this open until it is done.

[kenji-miyake]

I'll create a revert PR and relate this issue.

[ambroise-arm]

CUDA shouldn't be required to build Autoware

I believe it's not true.

I understand that cuda is necessary to have the runtime functionality of Autoware. But I strongly believe that it should be possible to built without it.
For example

autoware.universe/perception/lidar_apollo_instance_segmentation/CMakeLists.txt

Line 12 in 03de605

if(CUDA_FOUND)

has a different behavior whether cuda is present on the system or not, and doesn't fail if it's not.

I'll create a revert PR and relate this issue.

What about the problems from #1669, won't that just get us back to that?

EDIT: sorry, I hadn't seen #2410 (comment)

[kenji-miyake]

But I strongly believe that it should be possible to built without it.

Yes, it's true! We gave up ensuring no-cuda just due to CI jams. Let's re-enable it.