Offcreen framebuffer is not complete

[harjatinsingh]

Using offscreen rendering throws this error message with Tesla P100 GPU
ERROR: Offscreen framebuffer is not complete, error 0x8cdd.

However the same code works with 1080 Ti and Titan X GPUs. Is there something I need to modify. I am using mujoco1.5 with Ananconda Python 3.6.5.

[Gaoee]

I got a similar error too.

[ajabri]

Bump on this. Did you guys figure this issue out?

[yusukeurakami]

I have it too. In my case, it happens after I started using pytorch.multiprocess. Mujoco 1.55, Mujoco-py 1.5 Titan X GPU. Ubuntu 16.04.

I got following both errors.
ERROR: Shadow framebuffer is not complete, error 0x8cdd
ERROR: Offscreen framebuffer is not complete, error 0x8cdd

[btickell]

@yusukeurakami any luck?

[yusukeurakami]

@btickell
So, in my case, this problem happened because of the off-screen framebuffer became full by opening too many Mjviewers. Every iteration, I was creating the new MujocoEnv and calling MujocoEnv._get_viewer() for rendering purpose. However, I realized that the viewer will not be deleted even I deleted the env itself and eventually it will full up my off-screen memory space.
Therefore, I stopped calling the _get_viewer() function and use Unity to render the image instead of Mujoco renderer. Then I never got this error again.

[btickell]

Interesting, I am actually getting this error when I STOP naively generating new render contexts. In my case I am using offscreen rendering and when creating a new simulator I first access the original doing

rc = old_sim.render_contexts[0]
new_sim = mujoco_py.MjSim(model)
sim.add_render_context(rc)
rc.update_sim(new_sim)

If I remove this process the error goes away...

[tkelestemur]

I'm getting the same issue with the following:

    def __init__(self, render=False):

        self.model = mj.load_model_from_path(MODEL_PATH)
        self.sim = mj.MjSim(self.model)
        if render:
            self.viewer = mj.MjViewer(self.sim)
        else:
            self.depth_viewer = mj.MjRenderContextOffscreen(self.sim)

If the render is False it imideatly throws the same error. I don't even need to call or render or read_pixel functions. I'm on Ubuntu 18.04 with Titan RTX.

Edit: Exact same code works with Macbook Pro. So the issue is related to Ubuntu/nvidia.
To reproduce the error: https://github.com/tkelestemur/diff_histogram/blob/mocap/env/test_env.py

Edit 2: The read_pixels_depth functions works on both Mac and Ubuntu 18.04.
https://github.com/openai/mujoco-py/blob/master/mujoco_py/tests/test_opengl_context.py#L52