box3d_overlap may cause “RuntimeError: CUDA error: an illegal memory access was encountered” in some cases

[mxh1999]

🐛 Bugs / Unexpected behaviors

When I use box3d_overlap on GPU, it may cause "RuntimeError: CUDA error: an illegal memory access was encountered" under certain inputs.

Instructions To Reproduce the Issue:

from pytorch3d.ops import box3d_overlap
import torch
boxes1 = torch.tensor([[ 0.0622, -0.0626,  0.9963],
        [-0.0565,  0.1136,  0.7310],
        [-0.6875,  0.4739,  1.2525],
        [-0.5689,  0.2977,  1.5178],
        [ 0.2553,  0.1737,  1.0667],
        [ 0.1367,  0.3498,  0.8015],
        [-0.4944,  0.7101,  1.3230],
        [-0.3757,  0.5339,  1.5883]]).reshape(1,8,3).cuda()
boxes2 = torch.tensor([[-0.1583,  0.5106,  1.8723],
        [-0.1583,  0.5106,  0.2837],
        [-0.1583,  0.2942,  0.2837],
        [-0.1583,  0.2942,  1.8723],
        [-0.5130,  0.5106,  1.8723],
        [-0.5130,  0.5106,  0.2837],
        [-0.5130,  0.2942,  0.2837],
        [-0.5130,  0.2942,  1.8723]]).reshape(1,8,3).cuda()

_, iou3d = box3d_overlap(boxes1, boxes2)
print(iou3d)

This code will cause the problem. But if I remove .cuda(), it works fine.

I try to visualize these two boxes using open3d and I don't see anything unusual.

My Environment

python                    3.8.19               h955ad1f_0
pytorch                   1.11.0          py3.8_cuda11.3_cudnn8.2.0_0
pytorch3d                 0.7.2                    pypi_0
scipy                     1.10.1                   pypi_0
torchaudio                0.11.0               py38_cu113 
torchvision               0.12.0               py38_cu113
numpy                     1.24.3           py38hf6e8229_1  
numpy-base                1.24.3           py38h060ed82_1

[AndreasLH]

Running into the exact same issue here. using

• torch version 2.1.0
• pytorch3d version 0.7.6
• cuda_11.8.r11.8

[manurare]

The problem comes from MAX_TRIS when allocating a new array here. If MAX_TRIS is changed from 100 to 200 then it works. I wonder if it can be dinamically allocated instead... The intersection shape can have arbitrary number of triangles.