Replace GLX with EGL for X11

[christian-rauch]

This PR replaces GLX by EGL for the X11 windowing and fixes potential issues like: #74 #242 #260. Example programs like SimpleDisplay work already.

Before continuing working on this: Will replacing GLX by EGL be potentially merged? This will remove the GLX dependency but adds and dependency on EGL (which is also required by Android).

EDIT: With NVIDIA drivers > 384 (390, 396), the window content does not update and stays black.

Fixes #782.

[stevenlovegrove]

This is great! If we can make this reliable enough, I would be very keen to merge. Especially if we can support the use case of a frame buffer without window.

[christian-rauch]

I posted the issue with the blank window with certain nvidia drivers with a minimal example at the nvidia forum: https://devtalk.nvidia.com/default/topic/1037901/x11-with-egl-produces-black-window-for-driver-versions-gt-384 . Since this issue also appears in the official mesa example, I somehow suspect that this is a driver issue.
The other issue that shared EGL contexts are not supported. I.e. at the moment the program simply exits when trying to access a context from another thread. This might be related to the X11+GLX context sharing issue #393.

When this works, going to windowless should be easy. We could even separate the EGL context creation into a class that is used by X11, Wayland and windowless.

[christian-rauch]

The X11+EGL window and context creation now works in a threaded environment.
However, I needed to add an new method RemoveCurrent to the WindowInterface to unset the current context. When starting a new thread, you need to call BindToContext() to set the context in this thread, and call RemoveCurrent() to unset it. I added a new example HelloPangolinThreads that demonstrates the procedure.
The threading patches also work on the previous X11+GLX version. I included them here because they are the only way I could make threading with X11+EGL working. But I can create another dedicated PR for the threading patches if requiered.

[stevenlovegrove]

So did you manage to work around the NVidia issue that you references?

[christian-rauch]

The nvidia driver issue is unrelated to Pangolin, it also appears for the minimal EGL example that I posted in the nvidia forum https://devtalk.nvidia.com/default/topic/1037901/x11-with-egl-produces-black-window-for-driver-versions-gt-384 .

It seems to only occur with the nvidia drivers > 384 from the official CUDA repository for Ubuntu. E.g. it is working on other distributions or using the nvidia installation script (see post by matthewcmatl in the linked nvidia issue).
I have not tested this issue in many other configurations. It would be nice if someone with Ubuntu 18.04 could test if this occures with the drivers from the official package repo and the nvidia CUDA repo.

[stevenlovegrove]

Okay, I'm excited to merge this, but I guess I should wait until the Ubuntu package is fixed.

[christian-rauch]

I can confirm that the X11+EGL feature is working now with nvidia driver versions 390.48 (Ubuntu 18.04 repo) and 410.48 (CUDA repo).

[stevenlovegrove]

Are there any downsides to this implementation, or should I now consider this for merge?

[christian-rauch]

I am using this branch for quite a while now on Ubuntu 14.04 and 18.04 with Nvidia drivers from the CUDA repo. I think it would be good if this branch could be tested by others in different configurations to see if this issue with black windows still exists.

[fwcore]

Hi all, I tried this PR in a Ubuntu 22.04 docker with X11. HelloPangolin works as expected.

This PR helps me to solve my issue (libtorch+Pangolin segfault, see #884). Thank you very much.

The only thing I think is a little bit verbose is that I have to link explicitly OpenGL::EGL in CMakeLists.txt. I expect to link it implicitly through linking to pango_opengl, as for ${GLEW_LIBRARY}. See

https://github.com/christian-rauch/Pangolin/blob/5bf16e811704f98e12393f2a4ac5f5918e1f8fff/components/pango_opengl/CMakeLists.txt#L57C51-L57C66

Here is my current CMakeLists.txt

find_package(OpenGL REQUIRED COMPONENTS OpenGL EGL)
target_link_libraries(pangolin_libtorch_test
  PUBLIC
  "${TORCH_LIBRARIES}"
  ${Pangolin_LIBRARIES}
  OpenGL::EGL)

The EGL is explicitly added here. If I don't do this, I will have this error:

CMake Error at CMakeLists.txt:33 (add_executable):
  Target "pangolin_libtorch_test" links to target "OpenGL::EGL" but the
  target was not found.  Perhaps a find_package() call is missing for an
  IMPORTED target, or an ALIAS target is missing?

[EXing]

Thanks to this MR, save me from OpenGL errors

[christian-rauch]

@stevenlovegrove I've tested this successfully on Intel and Nvidia systems natively on X11 and via XWayland Wayland display servers. Since this removes the GLX dependency completely and fixes issues reported by others (#782, ), I would finally like to merge this now.

[stevenlovegrove]

Great, thanks!