EGL initialization error

[ajabri]

Hi,

I've recently come across the following error, on a machine with Nvidia driver version: 440.33.01 CUDA version: 10.2, when trying to use EGL for headless rendering. The same exact code was running properly on a machine with CUDA Version 10.1 and driver version 390.

Any idea how to fix the issue? I've tried reinstalling dm_control with pip.

  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/suite/__init__.py", line 28, in <module>
    from dm_control.suite import acrobot
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/suite/acrobot.py", line 24, in <module>
    from dm_control import mujoco
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/mujoco/__init__.py", line 18, in <module>
    from dm_control.mujoco.engine import action_spec
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/mujoco/engine.py", line 44, in <module>
    from dm_control import _render
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/__init__.py", line 75, in <module>
    Renderer = import_func()
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/__init__.py", line 36, in _import_egl
    from dm_control._render.pyopengl.egl_renderer import EGLContext
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_renderer.py", line 66, in <module>
    EGL_DISPLAY = create_initialized_headless_egl_display()
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_renderer.py", line 49, in create_initialized_headless_egl_display
    for device in EGL.eglQueryDevicesEXT():
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_ext.py", line 65, in eglQueryDevicesEXT
    success = _eglQueryDevicesEXT(max_devices, devices, num_devices)
ctypes.ArgumentError: argument 2: <class 'TypeError'>: expected LP_c_void_p instance instead of EGLDeviceEXT_pointer_Array_10

Thanks,
A

[xinleipan]

Try to do

$ export DISPLAY=:0

in your terminal and try again, it works for me.

[ajabri]

@xinleipan In this case, I'm trying to use EGL for headless rendering. Setting "DISPLAY=:0" results in the below error. Are you rendering with glfw or osmesa?

Traceback (most recent call last):
  File "main_maw.py", line 6, in <module>
    import dmc_wrapper as dmc2gym
  File "/home/aj/rlpyt/selfish/dmc_wrapper.py", line 2, in <module>
    from dm_control import suite
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/suite/__init__.py", line 28, in <module>
    from dm_control.suite import acrobot
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/suite/acrobot.py", line 24, in <module>
    from dm_control import mujoco
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/mujoco/__init__.py", line 18, in <module>
    from dm_control.mujoco.engine import action_spec
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/mujoco/engine.py", line 44, in <module>
    from dm_control import _render
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/__init__.py", line 67, in <module>
    Renderer = import_func()  # pylint: disable=invalid-name
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/__init__.py", line 36, in _import_egl
    from dm_control._render.pyopengl.egl_renderer import EGLContext
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_renderer.py", line 68, in <module>
    raise ImportError('Cannot initialize a headless EGL display.')
ImportError: Cannot initialize a headless EGL display.

[xinleipan]

I'm not sure, sorry. Maybe it's GLFW, I just installed MuJoCo normally with a Ubuntu 16.04 machine, CUDA 8.0 and NVIDIA GPU.

[geyang]

same issue here on a ubuntu box.

[geyang]

Problem solved, use this: @ajabri

export MUJOCO_GL="glfw"

In order to avoid the EGL, one way I have seen working is to use the GLFW backend. To do so, you need to set the environment flag to glfw. Then you need a display port that is actually available.

[ajabri]

I need headless rendering, so I need to use EGL as my rendering backend. The same code worked on another machine, the only difference being (I believe) that the machine on which I obtain the above error has a much more recent nvidia driver (440 vs 390).

[geyang]

@ajabri if X display is not available, you can use "osmesa" backend.

(for the interest of others: won't work even if you install Xvfb and GLFW because they require X-display)

export MUJOCO_GL="osmesa"

Here is the full setup for a headless linux box with no built-in X display:

      export MUJOCO_GL=osmesa
      export MJLIB_PATH=$HOME/.mujoco/mujoco200/bin/libmujoco200.so
      export MJKEY_PATH=$HOME/.mujoco/mujoco200/mjkey.txt
      export LD_LIBRARY_PATH=$HOME/.mujoco/mujoco200/bin:$LD_LIBRARY_PATH
      export MUJOCO_PY_MJPRO_PATH=$HOME/.mujoco/mujoco200/
      export MUJOCO_PY_MJKEY_PATH=$HOME/.mujoco/mujoco200/mjkey.txt

@saran-t maybe we can add this to the documentation? :)

[ajabri]

Hmmm, EGL should allow for hardware accelerated headless rendering without X display. Has worked properly for me in the past. Will look into other differences of current setup.

[geyang]

Maybe we can ping @saran-t about that

[lich14]

@ajabri if X display is not available, you can use "osmesa" backend.

(for the interest of others: won't work even if you install Xvfb and GLFW because they require X-display)

export MUJOCO_GL="osmesa"

Here is the full setup for a headless linux box with no built-in X display:

      export MUJOCO_GL=osmesa
      export MJLIB_PATH=$HOME/.mujoco/mujoco200/bin/libmujoco200.so
      export MJKEY_PATH=$HOME/.mujoco/mujoco200/mjkey.txt
      export LD_LIBRARY_PATH=$HOME/.mujoco/mujoco200/bin:$LD_LIBRARY_PATH
      export MUJOCO_PY_MJPRO_PATH=$HOME/.mujoco/mujoco200/
      export MUJOCO_PY_MJKEY_PATH=$HOME/.mujoco/mujoco200/mjkey.txt

@saran-t maybe we can add this to the documentation? :)

This way exactly solve me from those terrible mistakes

[saran-t]

I can't see how a CUDA upgrade could have changed anything given that we are specifying the argument types ourselves, but one thing that you could try is to change this line

https://github.com/deepmind/dm_control/blob/master/dm_control/_render/pyopengl/egl_ext.py#L65

to

success = _eglQueryDevicesEXT(max_devices, ctypes.POINTER(ctypes.c_void_p)(ctypes.addressof(devices)), num_devices)

Let me know if this works.

[ajabri]

The above change gives me the following error:

Traceback (most recent call last):
  File "main_maw.py", line 6, in <module>
    import dmc_wrapper as dmc2gym
  File "/home/aj/rlpyt/selfish/dmc_wrapper.py", line 2, in <module>
    from dm_control import suite
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/suite/__init__.py", line 28, in <module>
    from dm_control.suite import acrobot
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/suite/acrobot.py", line 24, in <module>
    from dm_control import mujoco
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/mujoco/__init__.py", line 18, in <module>
    from dm_control.mujoco.engine import action_spec
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/mujoco/engine.py", line 44, in <module>
    from dm_control import _render
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/__init__.py", line 67, in <module>
    Renderer = import_func()  # pylint: disable=invalid-name
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/__init__.py", line 36, in _import_egl
    from dm_control._render.pyopengl.egl_renderer import EGLContext
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_renderer.py", line 66, in <module>
    EGL_DISPLAY = create_initialized_headless_egl_display()
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_renderer.py", line 49, in create_initialized_headless_egl_display
    for device in EGL.eglQueryDevicesEXT():
  File "/home/aj/miniconda3/envs/rlpyt/lib/python3.7/site-packages/dm_control/_render/pyopengl/egl_ext.py", line 66, in eglQueryDevicesEXT
    success = _eglQueryDevicesEXT(max_devices, ctypes.POINTER(ctypes.c_void_p)(ctypes.addressof(devices)), num_devices)
TypeError: expected c_void_p instead of int

[alimuldal]

I think this was due to an upstream change in PyOpenGL. In versions newer than 3.1.4, OpenGL.EGL has its own EGLDeviceEXT member (mcfletch/pyopengl@38f4cc5), which overrides our assignment of EGLDeviceEXT = ctypes.c_void_p when we do the wildcard import on the last line. Consequently we end up with mismatching pointer types in _eglQueryDevicesEXT and in the body of eglQueryDevicesEXT. I'll put together a fix.

[ajabri]

Thank you @alimuldal!

[LazerLikeFocus]

Problem solved, use this: @ajabri

export MUJOCO_GL="glfw"

In order to avoid the EGL, one way I have seen working is to use the GLFW backend. To do so, you need to set the environment flag to glfw. Then you need a display port that is actually available.

I am new to this.
Can u plz tell me how can i use this 'export' function in colab?

[bdtacchi]

So I just spent hours on this, and I'm not an expert, so I might say something wrong. But, in my case, I was trying to run it over SSH, so I suspect that's what caused the headless display to not work. So, I tried switching the backend, but export MUJOCO_GL="glfw" or osmesa or whatever didn't make it not choose egl, so I had to go to the __init.py__ file from dmcontrol/_render and set the BACKEND variable right at the beginning to 'osmesa', which then made it finally use osmesa as the backend.