How to merge a large cuda array? about enoki

[merlinND]

Depending on how you structure your computation, you may have to do both. For example,

1. First do all per-sample computation in parallel, storing the per sample computation in a 600x800x32 array
2. Then accumulate and average the result to the image, e.g. using enoki::scatter_add

You could also decide to split your computation differently and repeat step (1) for each spp, then you would never have to create the larger array. In general, more parallelism is better, except when you run out of memory.

[andyyankai]

My code
c++:

python:

img = r.run()
print("done")
img = img.numpy() # this step can take a long long time if spp is high

However, this seems not a very efficient way to return the result. Is there any suggestion on how to fix this problem and optimize the code?