Enabling CUDA support for conda download

[Gadi Trocki]

Hello everyone, 

I am trying to run colmap on a cluster (running linux) using a script. The script runs fine for about half of it until it reaches the dense reconstruction. There, it fails giving the following error:

ERROR: Dense stereo reconstruction requires CUDA, which is not available on your system.

I know for a fact that the CUDAs are available in the system however, and I believe there is something wrong with the installation that doesn't support for the use of CUDA. 

I have also seen this in the documentation for the installation of colmap:

"Note that the COLMAP packages in the default repositories for Linux/Unix/BSD do not come with CUDA support, which requires manual compilation but is relatively easy on these platforms."

I don't know however how to do this manual compilation of the CUDA support to which they refer. I simply installed colmap from my conda environment using :

conda install -c conda-forge colmap

I would really appreaciate any help in enabling the CUDA support. 

Thank you very much in advance!

Gadi

[Damian Kopp]

Hello Gadi,

Since I am in the exact same situation as you were, I wonder if you figured it out and still remember how to fix the problem.
I would really appreciate help, thanks.

Damian

[Gadi Trocki]

Hello Damian,

I think i could in the end run part of the process in the gpu but had to use cpus for the patch match (or whatever it was called) which is the slowest part of the process. Bottom line is i gave up on using the cluster and simply run the reconstructions locally on a pc. Wish i could help more.

If you end up figuring it out please let me know how haha.

Good luck!

Gadi

--
You received this message because you are subscribed to a topic in the Google Groups "COLMAP" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/colmap/cWHy4B-eBuk/unsubscribe.
To unsubscribe from this group and all its topics, send an email to colmap+un...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/colmap/9ed2ec4c-058b-4366-9b48-f28047bb59d0n%40googlegroups.com.

[Ashish Doshi]

Hello Gadi,

On a cluster, usually you will need to load the CUDA drivers/toolkit first. Typically, this would `module load cuda`. Sometimes, you may have to specify cuda version as well such as `cuda/11.4`. However, I suggest you refer to your cluster documentation for more information. Then compile colmap. I can't remember the exact build configuration, but if the compiler finds cuda in the default location, then it will build colmap with cuda support (which would then let you run GPU based modules), otherwise it will run using CPU modules.

Another option is to build a Singularity container (using colmap's Dockerfile as reference as to what to include in your container) , as most HPC clusters do not allow for Docker containers to run.

Kind regards,

Ash.

[Gadi Trocki]

Hello Ash, 

Thanks for your response. I might need to use the cluster in the future so I'll keep the suggestions in mind. 

Best, 

Gadi

To view this discussion on the web visit https://groups.google.com/d/msgid/colmap/f0bf699e-32b5-4c73-ac14-c9fb90669eaan%40googlegroups.com.