Performance optimization for Hashgrid and RFTracer

NVIDIAGameWorks / kaolin-wisp

NVIDIA Kaolin Wisp is a PyTorch library powered by NVIDIA Kaolin Core to work with neural fields (including NeRFs, NGLOD, instant-ngp and VQAD).

Other

1.45k stars 132 forks source link

Closed hXl3s closed 1 year ago

hXl3s commented 1 year ago

Performance optimization PR:

Use packed representation of hashgrid codebook to reduce overhead of casting to FP16 in AMP training
Precompute masks for point filtering to reduce CPU <-> GPU synchronization count
Optimizer sets grads to None because it is faster than zeroing grads