search

openphotogrammetry / meshroomcl

MeshroomCL: An OpenCL implementation of photogrammetry with the Meshroom interface
Other
169 stars 8 forks source link

Discussion: What will it take the `Instant` NeRF to be imported to work with 'CL'? #44

Closed FatCache closed 2 years ago

FatCache commented 2 years ago

It is not related to meshroomcl project directly but I figured the best place to ask this question would be here considering folks behind this project managed to port entire Meshroom project to OpenCL!

I recently came across Nvidia's implementation of 'NeRF' and they managed to make it run in an instant mode. As far as I can tell, it is their implementation of https://github.com/bmild/nerf what has been optimized to run on Nvidia GPU - taking advantage of CUDA core/CUDA programming.

My question is, can this be ported to work in 'OpenCL' like it has done with meshroom? Or is there some fundamental piece of the puzzle missing here.

revisionarian commented 2 years ago

Hi @FatCache, great question! You are referring to NVIDIA's NGP Instant NeRF (https://github.com/NVlabs/instant-ngp) which is implemented with CUDA, and which also relies on their tiny-cuda-nn library that is CUDA-dependent.

Yes, in theory these two software components could be ported from CUDA to OpenCL. Our team does not have any immediate plans to port NeRF to OpenCL, but we have a lot of experience porting computer vision and graphics applications from CUDA to OpenCL, and we are following advances in neural rendering closely. If a good business case presents itself, then we would be interested in pursuing a porting effort.