docs(how-to guides): add runninng Autoware without CUDA

[kminoda]

Signed-off-by: kminoda koji.m.minoda@gmail.com

Description

It is better to explicitly mention that Autoware can be built and executed without CUDA (and also that there are some drawbacks).
Thus, I created a new page in "how-to guides" to explain how to do so.

Related link: https://github.com/orgs/autowarefoundation/discussions/209

Pre-review checklist for the PR author

The PR author must check the checkboxes below when creating the PR.

• I've confirmed the contribution guidelines.
• The PR follows the pull request guidelines.

In-review checklist for the PR reviewers

The Reviewers must check the checkboxes below before approval.

• The PR follows the pull request guidelines.

Post-review checklist for the PR author

The PR author must check the checkboxes below before merging.

• There are no open discussions or they are tracked via tickets.

After all checkboxes are checked, anyone who has write access can merge the PR.

[github-actions]

Documentation URL: https://autowarefoundation.github.io/autoware-documentation/pr-128/
Modified URLs:

• https://autowarefoundation.github.io/autoware-documentation/pr-128/how-to-guides/
• https://autowarefoundation.github.io/autoware-documentation/pr-128/how-to-guides/running-autoware-without-cuda/

[ghost]

@kminoda I have never tried to build Autoware without CUDA installed so I wanted to just check. Is CUDA only required when running Autoware, or do any errors/warnings occur when building any perception modules that use CUDA?

[kminoda]

@LalithVipulananthan
Some Autoware packages that depend on CUDA will not be built if CUDA is not available in the environment, for example, controlled in this line in CMakeLists.txt. So, CUDA is not necessary in the building process.

In running Autoware, launching CUDA-dependent packages (e.g. lidar-based detection) would fail if CUDA is not available.

[kminoda]

@LalithVipulananthan @kenji-miyake @soblin @shmpwk
Do you think we should write a concrete procedure for how to use a CUDA-free algorithm, especially for euclidean_cluster?

I personally feel that it is controversial, as the concrete procedure may change frequently due to the update around autoware_launch, tier4_*_launch, or euclidean_cluster and it is troublesome to keep this page up-to-date.

[shmpwk]

we should write a concrete procedure

No i don't think so.
IMO, simple instruction is enough because there is no cuda-free code option now.
The new users will feel helpful just knowing whether or not they can use Autoware without cuda.

[ghost]

@kminoda Apologies for the delay. I will review this PR by the end of today.

[ghost]

Reviewed and suggestions added.

[ghost]

Reworded/rephrased. Is it correct to say that the two lidar* modules related to LiDAR-based detection whilst tensorrt_yolo is used for camera-based detection? If so, then is it also correct to say that euclidean_cluster can onyl be used instead of the lidar-related modules?

Suggested change

	Instead of these modules, you can use the `euclidean_cluster` module, which is a 3D object detection algorithm that does not require a CUDA environment (refer to [the readme](https://github.com/autowarefoundation/autoware.universe/tree/main/perception/euclidean_cluster) for detail).
	Instead of these modules, you can use the `euclidean_cluster` module which contains a 3D object detection algorithm that does not require CUDA. For more details, refer to the [`euclidean_cluster` module's README file](https://github.com/autowarefoundation/autoware.universe/tree/main/perception/euclidean_cluster).

[kminoda]

Thank you for the suggestion!
I have reflected the comments at here 2f9ea05

Is it correct to say that the two lidar* modules related to LiDAR-based detection whilst tensorrt_yolo is used for camera-based detection?

Yes, you are correct.

If so, then is it also correct to say that euclidean_cluster can onyl be used instead of the lidar-related modules?

The current possible ways to perform 3D object detection are the following five configurations:

• lidar-apollo(-only)
• lidar-centerpoint(-only)
• lidar-apollo + tensorrt_yolo
• lidar-centerpoint + tensorrt_yolo
• euclidean_cluster(-only)

So, IMO the statement euclidean_cluster can only be used instead of the lidar-related modules is "partially" true?

[ghost]

@kminoda I think it is useful to know the different possible configurations so I have included them in my suggestion below. Alternatively, would it be better to include the information about possible configurations in the design documentation for the Perception component?

Suggested change

	The following three object detection modules require CUDA:
	- `lidar_centerpoint`
	- `lidar_apollo_instance_segmentation`
	- `tensorrt_yolo`
	Instead of these modules, you can use the `euclidean_cluster` module which contains a 3D object detection algorithm that does not require CUDA. For more details, refer to the [`euclidean_cluster` module's README file](https://github.com/autowarefoundation/autoware.universe/tree/main/perception/euclidean_cluster).
	Autoware Universe's object detection can be run using one of five possible configurations:
	- `lidar_centerpoint`
	- `lidar_apollo_instance_segmentation`
	- `lidar-apollo` + `tensorrt_yolo`
	- `lidar-centerpoint` + `tensorrt_yolo`
	- `euclidean_cluster`
	Of these five configurations, only the last one (`euclidean_cluster`) can be run without CUDA. For more details, refer to the [`euclidean_cluster` module's README file](https://github.com/autowarefoundation/autoware.universe/tree/main/perception/euclidean_cluster).

[ghost]

LGTM!