Create template for adding nvidia support to docker images

[ruffsl]

Context:

• Original: Changes for nvidia-docker v2 support car_demo#31
• Followup: opengl issure docker_images#105

Supplementary Context:

• Nvidia Docker Issue: nvidia-docker 1 can run OpenGL applications; nvidia-docker 2 can't NVIDIA/nvidia-docker#534
• ROS Wiki: Using Hardware Acceleration with Docker
• Nvidia's Openl Image: https://gitlab.com/nvidia/opengl/blob/ubuntu16.04/1.0-glvnd/devel/Dockerfile

Solution:

Perhaps we could add a template for adding the necessary configuration steps for supporting nvidia hardware acceleration in containers for users display forwarding rviz, gzviewer, or other opengl dependent programs. This could be done by mimicking nvidia's own image build steps. Then we could simply add a new tags for the OSRF docker repos that use the template for new child images.

Here is a rough but working example:

FROM osrf/ros:kinetic-desktop-full

RUN apt-get update && apt-get install -y --no-install-recommends \
        pkg-config \
        libxau-dev \
        libxdmcp-dev \
        libxcb1-dev \
        libxext-dev \
        libx11-dev && \
    rm -rf /var/lib/apt/lists/*

COPY --from=nvidia/opengl:1.0-glvnd-runtime-ubuntu16.04 \
  /usr/local/lib/x86_64-linux-gnu \
  /usr/local/lib/x86_64-linux-gnu

COPY --from=nvidia/opengl:1.0-glvnd-runtime-ubuntu16.04 \
  /usr/local/share/glvnd/egl_vendor.d/10_nvidia.json \
  /usr/local/share/glvnd/egl_vendor.d/10_nvidia.json

RUN echo '/usr/local/lib/x86_64-linux-gnu' >> /etc/ld.so.conf.d/glvnd.conf && \
    ldconfig && \
    echo '/usr/local/$LIB/libGL.so.1' >> /etc/ld.so.preload && \
    echo '/usr/local/$LIB/libEGL.so.1' >> /etc/ld.so.preload

# nvidia-container-runtime
ENV NVIDIA_VISIBLE_DEVICES \
    ${NVIDIA_VISIBLE_DEVICES:-all}
ENV NVIDIA_DRIVER_CAPABILITIES \
    ${NVIDIA_DRIVER_CAPABILITIES:+$NVIDIA_DRIVER_CAPABILITIES,}graphics

using --runtime=nvidia and existing X11 forwarding methods.
Perhaps @flx42 would have a better recommendation for something cleaner.

Personally, I find the necessity for multi-stage builds and assortment of in image copies not as elegant at mounting such files from the host and setting a three lightweight variables in the child dockerfile, as originally described in our ROS wiki:

FROM osrf/ros:indigo-desktop-full
# nvidia-docker hooks
LABEL com.nvidia.volumes.needed="nvidia_driver"
ENV PATH /usr/local/nvidia/bin:${PATH}
ENV LD_LIBRARY_PATH /usr/local/nvidia/lib:/usr/local/nvidia/lib64:${LD_LIBRARY_PATH}

But as nvidia docker v1 is deprecated and will become harder and harder to build from old sources, it would be nice to adopt a supported solution.

[tfoote]

I agree that the multi-stage build seems fragile and seems to break the idea of things being non-reproducible. From what I can see it's just installing some libraries and a json file it seems like it might be simpler for us to just build them in a normal build/ release install pipeline. If they are reproducible that shouldn't be a problem. Hopefully by doing that we're not limited to the only 2 architectures they appear to be targeting?

[flx42]

Looks like ubuntu 18.04 will provide packages for libglvnd, which would significantly simplify the dockerfile.
I think we should wait, since it's just a few weeks away.

[ruffsl]

I suppose libglvnd could be installed through the package manager when the docker image is based on 18.04, but I'm not sure that immediately helps ROS distros targeting 16.04 LTS. Perhaps its fare the newer nvidia docker v2 plugin is only supported with newer 18.04 based releases.

Which package is minimally necessary in this case: libglvnd-core-dev, libglvnd-dev, libglvnd0?
https://packages.ubuntu.com/search?keywords=libglvnd&searchon=names&suite=bionic&section=all

[flx42]

but I'm not sure that immediately helps ROS distros targeting 16.04 LTS. Perhaps its fare the newer nvidia docker v2 plugin is only supported with newer 18.04 based releases.

For 16.04 there are not many options: do something like you have above. Or use our own images for the base.
Or, but that's a marginal improvement, our OpenGL images are generic with an ARG for the FROM, which means you could be able to reuse the Dockerfiles to add libglvnd on top of your own images, with something like --build-arg from=osrf/...:
https://gitlab.com/nvidia/opengl/blob/ubuntu16.04/base/Dockerfile#L3
https://gitlab.com/nvidia/opengl/blob/ubuntu16.04/1.0-glvnd/runtime/Dockerfile#L42

Which package is minimally necessary in this case: libglvnd-core-dev, libglvnd-dev, libglvnd0?

libglvnd-dev and libglvnd0.

Other suggestions @3XX0?

[tfoote]

What about just a debian package that we inject to support libglvnd? Either a backport of the 18.04 package or manually build a package from the above procedure?

[ruffsl]

Just tried out GL Vendor-Neutral Dispatch library with nvidia-docker2 using 18.04. Works pretty slick:
NVIDIA/nvidia-docker#136 (comment)
Bump @flx42 .

[flx42]

Nice @ruffsl! Are you going to target only 18.04 then?

[ruffsl]

@flx42 , targeting melodic images on 18.04 using libglvnd for nvidia-docker2 sounds reasonable. As @tfoote mentioned however, would it be possible to backport libglvnd to say 16.04 so nvidia-docker2 could just easily be used on older mantined images? @tfoote , libglvnd appears to be available for debian stretch in backports.

https://packages.debian.org/search?keywords=libglvnd&searchon=names&suite=all&section=all

[tfoote]

We could backport it if there's not already a package for it available and it will let things work well on xenial. I see it's first available on artful: https://packages.ubuntu.com/source/artful/libglvnd

[abecciu]

+1 for 16.04 backport. Thanks for the hard work!

[flx42]

It's not handled by us, so I created a backport request:
https://bugs.launchpad.net/xenial-backports/+bug/1777944

[flx42]

I've been told a backport of this change for a LTS release carries risk, since you will need changes to the Mesa packages and our nvidia driver packages. Not saying it won't happen, but this use case is probably insufficient to convince them.

So, you should use ubuntu 18.04 or debian stretch I guess.

[ruffsl]

@tfoote , do you think it would make sense add the libglvnd install and ENVs for nvidia-docker to directly to desktop, or go the opposite route and create a new tag building off desktop-full, e.g. desktop-full-libglvnd, to append the needed layers. I'm more inclined to migrate it directly into melodic-desktop given libglvnd0 only adds about 25mb to that image (currently 2.05GB). I'm not aware of any downside though, as I'd expect most folks using the desktop image would be using them for the GUI tools they include, and if not, libglvnd wouldn't really changing things otherwise, but I'm not sure.

@mikaelarguedas , last time I think you modified the templates to white list distro specific changes:
8a83d5f . For this I'm thinking of just adding a new config here to include the additional packages or added image. Might be easier, but I wanted to check.

[tfoote]

Certainly adding the libglnvd0 for desktop and desktop-full and any other GUI specific targets makes sense to me.

[ruffsl]

I have some WIP here: osrf/docker_images#165
having some issues with rviz though.

[tfoote]

Now that we have the ability to run docker images in nvidia docker without baking in the nvidia before hand using: https://github.com/osrf/rocker

I think it makes sense to not bake those things into the core images.

[ruffsl]

@tfoote , perhaps you could reference rocker in the docker GUI entries in the ros wiki?