significantly downgraded performance when enabling `--ff` and `--cuda_ray` about torch-ngp HOT 3 CLOSED

@Karbo123 Hi, thanks for reporting this. It seems to be caused by bad hyperparameters in raymarching, which should be fixed in 873c267. You can try with the latest version!

from torch-ngp.

Yes, enabling --cuda_ray option in the latest commit (873c267) indeed increase the performance a bit. However, I also notice that if we enable --ff, the performance will downgrade slightly which is unexpected.
Here I give some test results.

CUDA_VISIBLE_DEVICES=2 python main_nerf.py data/fox --workspace trial_nerf --fp16 --ff --cuda_ray
==> 22 ~ 23 it/s

CUDA_VISIBLE_DEVICES=2 python main_nerf.py data/fox --workspace trial_nerf --fp16 --ff
==> 11 ~ 12 it/s

CUDA_VISIBLE_DEVICES=2 python main_nerf.py data/fox --workspace trial_nerf --fp16 
==> 22 ~ 23 it/s

CUDA_VISIBLE_DEVICES=2 python main_nerf.py data/fox --workspace trial_nerf 
==> 14 ~ 15 it/s

CUDA_VISIBLE_DEVICES=2 python main_nerf.py data/fox --workspace trial_nerf --cuda_ray
==> 30 ~ 33 it/s

from torch-ngp.

@Karbo123 On my TITAN RTX I got similar performance (16-17it/s) on fox with or without --ff.
This has been mentioned earlier #3 (comment), but unfortunately I don't have a 3090 to test.
Maybe you could just use the native --fp16, since it gives best performance.

from torch-ngp.