stage 0 mesh extraction RasterizeCudaContext segmentation fault

[SeryeongKim]

When executing the line self.glctx = dr.RasterizeGLContext(output_db=False) in the mark_unseen_triangles function in nerf/renderer.py during stage 0 mesh extraction,

Segmentation fault occurs
at self.cpp_wrapper = _get_plugin(gl=True).RasterizeGLStateWrapper(output_db, mode == 'automatic', cuda_device_idx)
in init() of class RasterizeGLContext in nvdiffrast/nvdiffrast/torch/ops.py.

I changed verbose=True option in torch.utils.cpp_extension.load in the _get_plugin method by referring to the nvdiffrast post. (NVlabs/nvdiffrast#46)

However, a segmentation fault occurs when running the RasterizeGLStateWrapper function.

Is there any way to solve this problem?

my docker environment is as follows

1. python3.8.10
2. cuda 11.6
3. cudnn 8
4. torch.version = 1.12.0+cu116

Thank you

[mowangmodi]

how to slove

[toomy0toons]

hi are you using a remote headless server?
i find its impossible to initialize a gl context with headless server.
try using cudarasterizer with max res 2048 and see if its works

[mowangmodi]

No,I use a remote head server

[SeryeongKim]

Using cudarasterizer not using raterizeglcontex with res 2048 is working. (like @toomy0toons simulation case)

But for the fine res (ex, higher than res 2048) simulation, raterizeglcontex should be used.

In my case, I should simulate with res >2048.

How can I solve segmentation fault by raterizeglcontex ?

[toomy0toons]

No,I use a remote head server
how are you getting display? if you are tunneling through like x11 or any remote methods you cannot use GL.
If not, check driver versions and softlinks to gl libs.

[toomy0toons]

Using cudarasterizer not using raterizeglcontex with res 2048 is working. (like @toomy0toons simulation case)

But for the fine res (ex, higher than res 2048) simulation, raterizeglcontex should be used.

In my case, I should simulate with res >2048.

How can I solve segmentation fault by raterizeglcontex ?

what is your gpu and nvidia driver version?
check if nvirasteizer is compiled with correct gpu . and you driver version supports it.
also it is impossible to render gl context if using remote headless server

[bgyarfas]

I had the same issue and it only happened when running the stage with --vis_pose.