Running time increase of PixelNeRFNet forward()?

[ty625911724]

During rendering, in the "composite()" in nerf.py, sample points will be split in to multiple batches, then run the PixelNeRFNet's forward() multiple times.

However, during tests, I found the running time of PixelNeRFNet forward() will increase during loops.
For example, in 10st loops, forward() just cost about 0.0015s. While in following loops, running time of forward() increase abruptly and cost about 0.18s.
This problem makes the rendering very slow. So, how can we avoid this undesirable increase?

Looking forward to your reply, thanks!