Illegal Memory Access using distCUDA2 (simple-knn) & Question

[azzarelli]

I was integrating the rasterizer (thank you) into the 4DGS pipeline and I came across an illegal memory access error that originated from initializing scales using from simple_knn._C import distCUDA2, i.e. initializing with:

def create_GS_from_point_cloud(...):
  ...
  dist2 = torch.clamp_min(distCUDA2(torch.from_numpy(np.asarray(pcd.points)).float().cuda()), 0.000001)
  scales = torch.log(torch.sqrt(dist2))[...,None].repeat(1, 3)

As a fix I implemented my own knn, but I'm curious whether this error originated from the simple-knn repo or this one.

Also I wanted to know whether anyone knows the main difference between the depth calculation here and the depth-diff-gaussian-rasterization method for depth?

[slothfulxtx]

As I remembered, there's no difference between our method for computing depth maps and depth-diff-gaussian-rasterizaition. Both of us apply alpha blending on the center depth value. I'm also curious about the oom error and how you fix it in your implementation. There are many reasons that may cause oom issue when computing distcuda2, hard to tell which is essential in your code.

[azzarelli]

I ended up replacing distcuda2 with this function, though I'm not totally sure if it's resolved the issues.

Wierdly, even just instantiating the simple-knn function, e.g. _ = distCUDA2(.) causes the error, so Im imagining it might be some conflict. I'm using the same Python and Cuda settings as you (3.11 and 11.8), so maybe you can replicate it? Might need to try this out on my work PC.

[slothfulxtx]

It seems that the original simple-knn repo is not reliable due to the later code updation. You can switch to other simple-knn repo URL to fix this problem