Possible to increase GPU usage?

[siddarth...@gmail.com]

Hi,

I'm looking into ways to accelerate the dense_stereo stage. I can start looking into multi-gpu solutions, but I noticed that colmap takes <1gb on my gpu, so I was wondering if there was anyway to allow it to parallelize further on the single gpu case.

[Johannes Schönberger]

There shouldn't be a benefit in running multiple tasks on the same GPU. But you can already do that now by simply passing the same GPU index multiple times to the ``--DenseStereo.gpu_index``. COLMAP also automatically uses multiple GPUs. Please refer to https://colmap.github.io/faq.html#multi-gpu-support-in-dense-reconstruction.

[s...@hivemapper.com]

Yes there doesn't seem to be any benefit. To be more specific, I was wondering if there was something analogous to increasing the block size in the feature matching stage ie. use more gpu memory to save on processing time if there was some hard-coded limit on how much gpu memory one could use. I am guessing that the bottleneck is somewhere else though.