Will Windows support be added, if so, when?

[Porfik]

[andybak]

Or failing that - I'd be interested to know if it's possible to run this under WSL with the recent support for direct GPU access.

[darglein]

I'm currently on an internship without access to a suitable windows machine. I will try it once I'm back in January.

[Ilyaprok]

Or failing that - I'd be interested to know if it's possible to run this under WSL with the recent support for direct GPU access.

Hello! I have tried in WSL2 with X server GUI, but unsuccessed. There is some render with 3D view. But neuro render "F2" closes app. Assertion 'cudaGraphicsGLRegisterImage(&graphic_resource, gl_buffer, gl_target, cudaGraphicsRegisterFlagsSurfaceLoadStore) == cudaSuccess' failed! File: /home/robot/neuro_render/ADOP/External/saiga/src/saiga/cuda/interop.h:72 Function: void Saiga::CUDA::Interop::initImage(unsigned int, gl::GLenum) Message: OS call failed or operation not supported on this OS Aborted My System: Windows 10 ver 21H2 19044.1288 WSL2 Ubuntu 20.04 ver 5.10.60.1-microsoft-standard-WSL2 Intel(R) Core(TM) i7-4700MQ CPU @ 2.40GHz 2.40 GHz 16 GB RAM x64 GeForce GTX860M Nvidia Driver 510.06 WSL2 CUDA ver 11.4

[darglein]

Does the libtorch/cuda stuff work on WSL? Can you try executing the test_loss and adop_train please?

If this works I can add an option to compile the viewer without CUDA/GL interop. This might reduce performance a little bit, but at least it displays an image.

[andybak]

In theory WSL should work with CUDA: https://docs.nvidia.com/cuda/wsl-user-guide/index.html

I think this works on Windows 11 without needing to run a preview build. Sadly I don't have either an preview build of Windows or a Windows 11 machine available right now as I decided to wipe a box to install Linux normally.

[robsingh9]

I tried on WSL2 for Windows 10 and could not get it working. The compile works fine until make -j10 where nothing happens. Works great on Ubuntu however.

[metya]

I tried on WSL2 too. And it is fully compilable, but doesn't work and doesn't render anything because of an error with syscalls in windows I think.

So I tried it in my Arch and it is doesn't compile at all.

So even if it is on Linux, it is doesn't working((

[darglein]

@metya What is the issue on Arch? Did you follow the install scripts?

[blurgyy]

So I tried it in my Arch and it is doesn't compile at all.

So even if it is on Linux, it is doesn't working((

That is not true, I have managed to compile on latest Arch after some tinkering. Some notable changes include:

1. Install cuda-11.1 and gcc-9 on your system (because in the AUR there is no cuda-11.2 and I find cuda-11.1 to work just fine with cudatoolkit=11.2), and prepend /opt/cuda-11.1 to your CUDA_PATH and some relevant paths to your PATH (refer to your /etc/profile.d/cuda.sh)
2. Install freeimage in your conda environment (this seems to be added since commit ec1432e)
3. Compile srt with the same environment as PyTorch and ADOP, otherwise compilation of ADOP fails at the final linking stage. You could first add https://github.com/Haivision/srt.git to submodules, then create a install_srt.sh similar to the install_pytorch.sh provided in the repo.

I'm not here to give support, but you should certainly try harder, especially when you decided to use Arch.

[darglein]

@blurgyy Thanks for the short Arch guide. May I ask why you need a system-wide install of CUDA? Shouldn't the cuda toolkit which is installed in the conda environment be enough?

[blurgyy]

May I ask why you need a system-wide install of CUDA? Shouldn't the cuda toolkit which is installed in the conda environment be enough?

Oh I think I just tried to match the environment given in the repo in the first place, did not try to build with latest CUDA (11.5 on Arch). You are probably right! (Also, I just checked that I also created my environment with cudatoolkit=11.1 to match a system CUDA version 11.1)

[metya]

@darglein Yes. I completely follow the install instructions and scripts row by row. since it is not very related to this issue, I create a new one.

I will try instruction from @blurgyy and if it doesn't help, I'll create a new issue and will link it here.

@blurgyy thanks! I will try it! I've installed freeimage in my conda env. But I've realized that I have cuda-11.4.2-1 in my arch. And didn't compile srt by own.

[metya]

So I tried what said @blurgyy and it didn't work. I described details here #27

[stjema]

i can share my screen. i have win11, we can see if it works