Is CUDA necessary for Autoware?

[kminoda]

Hi everyone.

Is CUDA or GPU necessary for Autoware?

--
Why I am asking so is because I encountered an error due to the nonexistence of GPU.

Currently, I am trying to build an autonomous driving mobile robot without GPU or CUDA.
Thus, the object detection package (both centerpoint and apollo_instance_segmentation) fails to launch with the following error:

[lidar_apollo_instance_segmentation_node-1] ERROR: CUDA initialization failure with error 100. Please check your CUDA installation:  http://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html

Due to this node's failure, there is no output from /perception/object_recognition/objects.
This ends up in a failure of a behavior planning node that is subscribing to the topic, possibly caused by here in behavior_path_planner package.

So, we cannot properly start behavior / path planning in autoware.launch.xml.

Is this an intended behavior?

If not, can you modify this? Also, should I open an issue for this?

(I think that one of the possible solutions for this would be to add a functionality in lidar_apollo_instance_segmentation and lidar_centerpoint to, when CUDA is not available, publish dummy empty messages to /perception/object_recognition/objects. How about this?)

[kenji-miyake]

It is explained here.

autoware/ansible/playbooks/universe.yaml

Line 12 in 5fd3659

[Warning] Some components depend on NVIDIA libraries (CUDA, cuDNN, TensorRT), which require license agreements.

autoware/ansible/playbooks/universe.yaml

Line 20 in 5fd3659

[Warning] Skipping installation of NVIDIA libraries. Please manually install them if you use the depending components.

So the answer is,

• Some Autoware "Universe" components currently depend on CUDA.
• It might be replaced with https://github.com/autowarefoundation/tvm_vendor in the future. (But not planned yet.)

when CUDA is not available, publish dummy empty messages to /perception/object_recognition/objects. How about this?

Yes, it's a good idea.
But it's not good to automatically do such a thing. Users should manually switch some flag to use dummy nodes.

[kminoda]

I see, thank you for the answer!

Yes, it's a good idea.
But it's not good to automatically do such a thing. Users should manually switch some flag to use dummy nodes.

Thanks!
Are there any other ways to solve this problem? For instance, does autoware have any packages that perform object detection from LiDAR without GPU?

[kminoda]

That's a huge help!! Thanks!

[shmpwk]

I hope in the future autoware can choose euclidean_cluster node instead apollo or centerpoint.
autowarefoundation/autoware.universe#842

[kenji-miyake]

Also, TVM transition is going, which I believe can remove CUDA dependency.
autowarefoundation/autoware.universe#908

[ghost]

In the meantime, would it be useful to update the Autoware documentation to be more explicit about what will happen if CUDA is not installed? Eg:

• List all of packages that will fail to build without CUDA installed (maybe this information should be added to the troubleshooting guide 🤔 )
• Explain that the euclidean_cluster node can be used instead of apollo or centerpoint for LiDAR based object detection (not sure where to put this information. Also in troubleshooting?)
  • Bonus: Provide explicit steps showing how to use the euclidean_cluster node instead of the default nodes.

[kminoda]

Agreed! I will create a new page for that: autowarefoundation/autoware-documentation#128