search

graphdeco-inria / nerfshop

NeRFshop: Interactive Editing of Neural Radiance Fields
https://repo-sam.inria.fr/fungraph/nerfshop/
Other
451 stars 24 forks source link

GPU architecture error when compling on windows #13

Closed samcao416 closed 1 year ago

samcao416 commented 1 year ago

I'm using a 4090 GPU and windows 11. When compling the cmake file in windows terminal after entering cmake --build build --config RelWithDebInfo -j, there comes an error, calling nvcc fatal : Unsupported gpu architecture 'compute_89'

However, I could successfully compile and build instant-ngp on my computer, before and after the failure building nerfshop.

Snosixtyboo commented 1 year ago

Hi,

thanks, that's an interesting error. Iirc, tinycudnn for InstantNGP does automatic detection of the latest possible compute capability for your GPU, which would be 89 for a 4090. But apparently, your installed CUDA toolkit is older and doesn't recognize this capability yet. Instant-NGP code has changed since we forked from it, getting nerfshop to behave the same might be rather tough. If it's possible, a potential fix would be to try and install the latest CUDA toolkit?

samcao416 commented 1 year ago

Thanks for your timely reply.

Yeah, I am using cuda 11.7.1. Previously, instant NGP didn't support the lastest CUDA. For stability, I only used CUDA 11.7.1. I will try the latest 12.1 and back to you soon.

samcao416 commented 1 year ago

After alternating to cuda 12.1, the compilation completed successfully. Thanks a lot!