Move CUDA-dependent packages to a separate repository

[esteve]

Checklist

• I've read the contribution guidelines.
• I've searched other issues and no duplicate issues were found.
• I've agreed with the maintainers that I can plan this task.

Description

As discussed in autowarefoundation/autoware#3222 (comment), in order to be able to build Autoware.Universe with the available packages in the ROS index, we decided to move those packages that depend on CUDA to a separate repository

Purpose

Be able to compile a subset of the Autoware.universe without depending on CUDA

Possible approaches

Move the the packages that depend on CUDA to a separate repository

Definition of done

There are no packages that depend on CUDA in Autoware.universe

[miursh]

Is there any options not to separate CUDA dependent packages?
Because of reasons below, I would not want to separate just a part of perception packages

• If some of perception packages are placed in the different repository, it is very confusing for developers.
  • For example, when we want to add some change, we might need to make 3-4 PRs to different repository, such as new perception repo, autoware.unvierse, autoware_launch, and perception utils (new repo too?)
• Very difficult to understand especially for new users
• There might be some dependency between autoware.universe and separated packages. If there is many dependency, github CI might not work well.

[xmfcx]

Name of the new cuda packages repository

What are your naming choices about this new repository?

I'd go for cuda_autoware_pkgs, what are your suggestions?

Make naming consistent

Also after moving the packages, we should make the naming conventions uniform too.

And for that my proposal is:

• cuda_trt_package_name
• cuda_package_name (we don't have any cuda only packages yet, just convention)
• cuda_tvm_package_name (this is also preserved for future, we will have a wide discussion on what to do with tvm packages in general soon)

Handling of TVM packages

Also we should move tvm packages to tvm_autoware_pkgs and they will contain tvm packages for non-cuda targets.

This topic will also be discussed by @ambroise-arm @angry-crab @mitsudome-r @esteve @kenji-miyake and me.

to separate or not to separate

Edit: @miursh @yukkysaito I've just saw your comment after refreshing. Yes, I will think about not separating too. I understand the concerns, this is an issue with all separated repositories.

[kenji-miyake]

I'm sorry for joining late.

To give you some additional information, splitting repositories helps us reduce CI time. There is no need to test all packages with CUDA.

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

Lines 15 to 17 in 541122f

	container-suffix:
	- ""
	- -cuda

But I understand splitting repositories may increase maintenance costs and troubles.

I think it's okay not to split repositories right now, we should do the following at least:

• As @xmfcx says, make package names consistent, which helps us ignore packages in CI.
  • Related: Is there a way to ignore some packages from being built? jspricke/ros-deb-builder-action#19
• Plan a migration timeline to be CUDA-independent for hardware-agnostic, which I believe is one of the goals of Autoware.

I'll discuss how to proceed with this also in TIER IV.

[esteve]

@kenji-miyake another reason for splitting the CUDA-related packages to another repository is because when/if we submit Autoware to the ROS buildfarm that Open Robotics maintains (https://build.ros.org/), all the packages in the repository are submitted and they need to be distributable and not covered by a proprietary license.

However, following @mitsudome-r 's rationale of only submitting Autoware.core and not all of Autoware.universe to the ROS buildfarm, we can either leave CUDA packages in universe, or move them in a separate repository, but we can't migrate them to Autoware.core.

[kenji-miyake]

However, following @mitsudome-r 's rationale of only submitting Autoware.core and not all of Autoware.universe to the ROS buildfarm, we can either leave CUDA packages in universe, or move them in a separate repository, but we can't migrate them to Autoware.core.

Yes, I think so, too.

[kenji-miyake]

As a result of TIER IV's internal discussion, we'd like to propose the following actions:

• To resolve the build issues in Debian package release, we'll rename existing packages, and don't move packages to another repository yet.
  • Having a unified naming convention can help us ignore specific packages in CI.
  • Moving packages to another repository has significant impacts, so we need step-by-step improvements.

For that,

• Define naming conventions.
  • For CUDA/TensorRT packages, use _trt.
  • For TVM packages, use _tvm.
• Rename existing packages to follow the defined conventions.
  • In addition, maybe old-style packages should be fixed.
    • For example, https://github.com/autowarefoundation/autoware.universe/tree/aba954f30124155afe15ba4f216076101a125265/perception/lidar_apollo_segmentation_tvm_nodes can be merged into https://github.com/autowarefoundation/autoware.universe/tree/aba954f30124155afe15ba4f216076101a125265/perception/lidar_apollo_segmentation_tvm.
• Write the naming convention rules in the contribution guidelines.

---

@xmfcx @esteve cc @mitsudome-r What do you think about this?
If it looks good, we'd appreciate it if @esteve -san would work on these tasks as part of the Debian package issue.

Regarding the long-term milestones, I think as follows:

• We should continue discussing how we should utilize TVM.
• Each one should try to gain a deeper understanding around TVM through learning/experiments.

[xmfcx]

I agree that we can start by renaming.

• For CUDA/TensorRT packages, use _trt.
• For TVM packages, use _tvm.

I'd prefer prefix over suffix. But I'm OK either way.

Also merging of _nodes double-packages will simplify the maintenance in general.

[esteve]

I agree that we can start by renaming.

• For CUDA/TensorRT packages, use _trt.
• For TVM packages, use _tvm.

I'd prefer prefix over suffix. But I'm OK either way.

I also agree with using suffixes, makes it clear that they are an extension to the base packages or that they are based on other packages.

Also merging of _nodes double-packages will simplify the maintenance in general.

I'm afraid I don't follow you here, can you elaborate? Thanks.

[xmfcx]

I'm afraid I don't follow you here, can you elaborate? Thanks.

For example,

• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/lidar_apollo_segmentation_tvm
• https://github.com/autowarefoundation/autoware.universe/tree/main/perception/lidar_apollo_segmentation_tvm_nodes

could be merged into a single package.

[esteve]

@xmfcx thanks for elaborating.

I can't clearly remember what was the rationale for having _nodes packages, I think it was inherited from Autoware.Auto because Apex.AI wanted the core functionality separate from the node code because it was incompatible with Apex.OS, but I might be making this up, I don't remember.

Anyway, I agree with you, it'd simplify our tree, especially now that we're enforcing all nodes to be components, so they all have uniform instantiation.

[xmfcx]

I can't clearly remember what was the rationale for having _nodes packages, I think it was inherited from Autoware.Auto because Apex.AI wanted the core functionality separate from the node code because it was incompatible with Apex.OS, but I might be making this up, I don't remember.

From what I remember, it was to separate ROS libraries from the rest of the packages. And it's a good practice in theory because the library can also be reused by other packages too. But in the implementation, this wasn't adhered strictly. So I think it lost its purpose along the way.

Found the source:
https://autowarefoundation.gitlab.io/autoware.auto/AutowareAuto/contributor-guidelines.html#autotoc_md24

This design pattern helps to promote separation of concerns and code re-use. The core and the ROS node(s) can be implemented in separate packages; e.g. foo and foo_nodes. There are some trivial cases where a simple ROS Node that does not require a "core" are acceptable but these should be the exception, not the rule.

Also I used to be 3-tier before (but I think it's just a matter of semantics for the executable):
https://gitlab.com/autowarefoundation/autoware.auto/AutowareAuto/-/merge_requests/255#note_316711298

[stale]

This pull request has been automatically marked as stale because it has not had recent activity.

[esteve]

@xmfcx @idorobotics @mitsudome-r sorry for reviving a stale issue, but it's not clear to me whether we reached an agreement here. I'm working on adding support for building CUDA-enabled packages in our buildfarm and it'd be good to sort this topic out. Thanks.

[xmfcx]

How to structure the separated CUDA packages. The reason to separate CUDA packages is because CUDA is hardware specific.

For both Universe and Core, 2 new repositories:

• autoware/src/universe/autoware.universe_cuda
• autoware/src/core/autoware.core_cuda (if necessary, in the future)

Additional:

• We should update the installation instructions as well for separated packages.
• Update the autoware.repos and autoware-nightly.repos files to have them.

Pros:

• Easier to submit Universe to build farm
• Easier to maintain CI
• Easier to generate debian packages

Cons:

• More work to maintain with Universe
  • Multiple PRs

@YoshiRi @miursh @youtalk @oguzkaganozt

In the future:

• CUDA dependent packages should fail if there is no CUDA.