Unable to use SYCL device in raycasting example

[fabiannagel]

Checklist

• I have searched for similar issues.
• For Python issues, I have tested with the latest development wheel.
• I have checked the release documentation and the latest documentation (for main branch).

Describe the issue

I've been looking into the new SYCL support in v0.19 (using this wheel) and managed to get everything set up to the point where o3d.core.sycl.print_sycl_devices(print_all=True) gives me a green light:

[Open3D INFO] # Open3D SYCL device
[Open3D INFO] - Device("SYCL:0"): [ext_oneapi_level_zero:gpu] Intel(R) UHD Graphics 770
[Open3D INFO] # All SYCL devices
[Open3D INFO] - [opencl:gpu] Intel(R) UHD Graphics 770
[Open3D INFO] - [ext_oneapi_level_zero:gpu] Intel(R) UHD Graphics 770
[Open3D INFO] # Default SYCL selectors
[Open3D INFO] - sycl::default_selector_v    : [ext_oneapi_level_zero:gpu] Intel(R) UHD Graphics 770
[Open3D INFO] - sycl::cpu_selector_v        : N/A
[Open3D INFO] - sycl::gpu_selector_v        : [ext_oneapi_level_zero:gpu] Intel(R) UHD Graphics 770
[Open3D INFO] - sycl::accelerator_selector_v: N/A

Also:

# o3d.core.sycl_demo()
[Open3D INFO] SYCLDemo passed!

# print(o3d.cpu.pybind.core.sycl.is_available())
True

However, I verified that the raycasting example runs on CPU (no GPU load in intel_gpu_top) by default. So I tried explicitly setting the device - which fails. What am I doing wrong here?

Steps to reproduce the bug

import open3d as o3d
import matplotlib.pyplot as plt


for i in range(1000):
    cube = o3d.t.geometry.TriangleMesh.from_legacy(
                                        o3d.geometry.TriangleMesh.create_box())

    device = Device("ext_oneapi_level_zero:gpu")
    # Create scene and add the cube mesh
    scene = o3d.t.geometry.RaycastingScene(device=device)
    scene.add_triangles(cube)

    # Rays are 6D vectors with origin and ray direction.
    # Here we use a helper function to create rays for a pinhole camera.
    rays = scene.create_rays_pinhole(fov_deg=60,
                                     center=[0.5,0.5,0.5],
                                     eye=[-1,-1,-1],
                                     up=[0,0,1],
                                     width_px=3200,
                                     height_px=2400)

    # Compute the ray intersections.
    ans = scene.cast_rays(rays)

    # Visualize the hit distance (depth)
    plt.imshow(ans['t_hit'].numpy())

Error message

RuntimeError: [Open3D Error] (void open3d::t::geometry::RaycastingScene::SYCLImpl::InitializeDevice()) /root/Open3D/cpp/open3d/t/geometry/RaycastingScene.cpp:459: Caught exception creating sycl::device: No device of requested type available. Please check https://software.intel.com/content/www/us/en/develop/articles/intel-oneapi-dpcpp-system-requirements.html -1 (PI_ERROR_DEVICE_NOT_FOUND)

Expected behavior

Open3D uses the SYCL:0 device and generates load on the GPU.

Open3D, Python and System information

- Operating system: Ubuntu 24.04
- Python version: Python 3.12.3
- Open3D version: output from python: 0.19.0
- System architecture: amd64
- Is this a remote workstation?: yes
- How did you install Open3D?: Remote wheel via `uv`

[ssheorey]

Hi @fabiannagel You need a GPU with HW raycasting support for this. I believe it's only the Arc / Xe series GPUs (discrete or integrated) that have this support. I'll update the docs to make this clear.

[fabiannagel]

Thanks @ssheorey, that makes sense. Is there any info on what operations are supported by the SYCL-backend?