GPU Memory leak!! about torch_kdtree HOT 5 CLOSED

Yeah, that is a problem, as we haven't checked the codes of cuda memory allocation yet. We will check the codes as soon as possible.

from torch_kdtree.

The cuda memory leakage problem is fixed in the latest commit. Test script is provided in test/test_memleak.py.
Further questions are welcome.

from torch_kdtree.

Thank you for your reply.
Btw， there are still two problem I found.

1. The tree can't be built when the number of points is equal of less than 2^14. I think the reason is that the number of threads and blocks is fixed. When I decrease the blocks from 32 to 16, it work well with 2^14 points.
2. And if I want to build a tree with N numbers while log2(N) is not a integer, it can't be built either. Maybe it's also caused by thread operation.

from torch_kdtree.

Thanks, and they are really some problems.
The codes of building kdtree is modified from another repo KdTreeGPU, and since KdTreeGPU is limited to those two problems you mentioned, so for now our torch_kdtree also suffers from these problems. This repo is still under development, functions could be enhanced and problems could be fixed in the future releases.

from torch_kdtree.

I got it. Thank you for your reply.

from torch_kdtree.