---

name: ops-metrics-conversions description: "Batching, representation conversions, SPC, Gaussian, quaternion, and metric workflows for Kaolin tensors." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

ops-metrics-conversions

Use this sub-skill for in-memory Kaolin work where geometry is already a tensor or Kaolin container and the task is about tensor batching, operations, conversions, math, or metrics.

Use when

• Choosing list, exact, packed, or padded layout for variable-size 3D tensors.
• Sampling/inspecting triangle meshes, point clouds, voxelgrids, SPC octrees, or Gaussian tensors.
• Converting among mesh, pointcloud, voxelgrid, SPC, and Gaussian occupancy representations.
• Applying kaolin.math.quat quaternion, rotation, Euclidean, or compact transform helpers.
• Computing pointcloud, point-to-mesh, voxelgrid, mask, or mesh-regularization metrics/losses.

Route elsewhere

• File I/O, datasets, SurfaceMesh construction, materials, USD/OBJ/PLY/GLTF/container selection: geometry-io-representations.
• Cameras, lighting, rasterization, differentiable rendering, ray tracing, images: rendering-cameras-lighting.
• Simplicits/Newton/Warp simulation loops and training-state physics: physics-simulation.
• Timelapse, Dash3D, notebooks, browser/WebGL display: visualization-workflows.

Read first

• API reference for function names, shape contracts, and backend gates.
• Workflows for layout selection, conversions, Gaussian transforms, quaternion recipes, and metric choices.
• SPC workflows for octree, lengths, pyramids, exsum, query/interpolation, and convolution conventions.
• Troubleshooting for packed/padded, CPU/CUDA, Warp, SPC lengths, and legacy/current exsum failures.

Bundled probes

• Quaternion smoke: CPU-safe quaternion/transform round trips.
• Tensor ops smoke: CPU tensor, mesh, voxel, Gaussian-transform, metric checks; optional --cuda-smoke for CUDA/SPC paths.

Operating rules

1. Keep layout explicit: list, exact batch, packed, padded, dense voxelgrid, sparse SPC, per-face feature, or per-point feature.
2. Prefer packed layouts when element counts vary and downstream code supports flattened indexing; use padded layouts only when dense batch axes are required.
3. Carry shape_per_tensor, numel_per_tensor, and first_idx with packed data.
4. For SPC scan_octrees, keep lengths on CPU and prefer the current exsum layout from scan_octrees(octrees, lengths).
5. kaolin.math.quat uses xyzw quaternions; transform_gaussians defaults to wxyz unless use_xyzw=True.
6. Gate CUDA-only operations explicitly; a CPU import is not proof that SPC, marching-cubes, Gaussian voxelization, or pointcloud metric kernels are available.