Open3D (v1.7)-ZHENWU PPU Cloud Service(ppu)-阿里云帮助中心

When you install certain 3D detection frameworks, such as mmdetection3d, Open3D is often installed automatically. However, pip install open3d installs a pre-compiled version that includes a GPU binary incompatible with PPU. You must rebuild it from source using the steps below.

The following steps use Open3D v0.18.0 as an example and have been tested and verified in a Docker environment with Torch 2.4 and CUDA 12.3.

Build steps

git clone -b v0.18.0 https://github.com/isl-org/Open3D
cd Open3D

apt update
apt install -y xorg-dev libxcb-shm0 libglu1-mesa-dev libc++-dev libc++abi-dev

# for cuda 12.4+
# If you use CUDA 12.4 or later, apply the community patch to upgrade the stdgpu version
git cherry-pick -X theirs b7e5f163e47213fb59eb93148ebd133360cbab26
# for cuda 11.x
touch /usr/local/PPU_SDK/CUDA_SDK/lib64/liblapack_static.a
# for torch 2.x
sed -i 's/CMAKE_CXX_STANDARD 14/CMAKE_CXX_STANDARD 17/g' CMakeLists.txt

mkdir build
cd build
cmake -DBUILD_CUDA_MODULE=ON \
      -DBUILD_PYTORCH_OPS=ON \
      -DCMAKE_BUILD_TYPE=RelWithDebInfo \
      ..
make -j$(nproc)
make install
pip install yapf
make pip-package

pip install lib/python_package/pip_package/*.whl

python -c "import open3d"

Unit test results

dir	file
.	test_color_map_optimization.py	1 passed
	test_flags.py	1 passed
	test_octree.py	9 passed
	test_open3d_eigen.py	39 passed
core	test_core.py	552 passed
	test_hashmap.py	16 passed
	test_linalg.py	2 failed, 82 passed
	test_nn.py	12 passed
	test_size_vector.py	3 passed
	test_sycl_utils.py	1 skipped
	test_tensor_function.py	40 passed
	test_torch_integration.py	5 passed
data	test_data.py	31 passed, 4 skipped
io	test_io.py rpc/test_remote_functions.py rpc/test_serialization.py	3 passed, 1 skipped
ml	test_contrib_iou.py	4 passed
ml	test_contrib_subsample.py	2 passed
ml_ops	test_cconv.py	8 passed, 3 skipped
	test_fixed_radius_search.py	3842 passed
	test_general_sparseconv.py	8 passed
	test_invert_neighbors_list.py	2 failed, 42 passed
	test_ragged_tensor.py	16 failed, 10 passed
	test_voxelize.py	2688 passed
	...	682 passed, 12 skipped
t	geometry/test_image.py	6 passed
	geometry/test_lineset.py	4 passed
	geometry/test_pointcloud.py	12 passed
	geometry/test_raycasting_scene.py	18 passed
	geometry/test_tensormap.py	9 passed
	geometry/test_trianglemesh.py	43 passed
	io/test_noise.py	1 passed
	io/test_realsense.py	2 skipped
	registration/test_registration.py	12 passed
	registration/test_transformation_estimation.py	8 passed
visualization	test_cpu_rendering.py	1 skipped
visualization	test_tensorboard_plugin.py	4 skipped

core/test_linalg.py fails because it uses the unsupported cusolverDnSgesvd_bufferSize and cusolverDnDgesvd_bufferSize functions.
A few test cases in the ml_ops directory result in a core dumped error at runtime.
The current pass rate is 99.7% (8193/8213).

Known issues

Unit test runtime issues
- test_linalg.py::test_svd
  - The open3d.core.svd operation fails because it uses the unsupported cusolverDnSgesvd_bufferSize and cusolverDnDgesvd_bufferSize functions.
- test_invert_neighbors_list.py
  - Passing an incorrectly shaped input to ml.ops.invert_neighbors_list causes a segmentation fault (core dumped) error instead of the expected Exception.
- test_ragged_tensor.py
  - Iterating through the elements of a RaggedTensor causes a segmentation fault (core dumped) error.