| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.SpaceToDepth |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 13 |
|
|
| ## Description |
|
|
| Rearranges blocks of spatial data into depth by moving values from the height and width dimensions into the channel dimension. An NCHW input of shape `[N, C, H, W]` produces an output of shape `[N, C * blocksize * blocksize, H / blocksize, W / blocksize]`. |
|
|
| See the [ONNX `SpaceToDepth` spec](https://onnx.ai/onnx/operators/onnx__SpaceToDepth.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `input` | `T` | `4` | — | 4-D input tensor of shape `[N, C, H, W]`. | required | |
|
|
| ## Outputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `output` | `T` | `4` | derived | 4-D output tensor of shape `[N, C * blocksize * blocksize, H / blocksize, W / blocksize]`. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `blocksize` | — | Size of the spatial block to collapse into depth; each `blocksize x blocksize` patch of pixels becomes additional channels. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16`, `int32`, `int16`, `int8`, `uint32`, `uint8`, `bool` | |
|
|
| ## Implementation variants |
|
|
| One implementation is selected per call from the device capabilities, the request shapes and the dtypes; these notes say what each one covers. |
|
|
| - `flat_vector` — Groups four flat outputs per invocation, including groups that cross spatial rows or channels. Reuses the shared permutation coordinates and writes one contiguous vector. |
| - `collapsed_input_vector` — When one spatial block spans the input plane, reads contiguous vec4 words and writes channel streams. Spatial tiles balance the available channel workgroups; paired f16 channels give each invocation whole output words. |
|
|
| ## Files |
|
|
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, per-variant templates, provenance) |
| - [`manifest.json`](build/webgpu/manifest.json) — the op contract (source of truth) |
| - [`test.json`](build/webgpu/test.json) — correctness cases |
| - [`bench.json`](build/webgpu/bench.json) — benchmark cases |
| - [`space-depth-collapsed-input.wgsl.jinja`](build/webgpu/space-depth-collapsed-input.wgsl.jinja) |
| - [`space-depth-permute.wgsl.jinja`](build/webgpu/space-depth-permute.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.3 |
| ``` |
|
|
| Required output shapes and logical data types are inferred from the supplied inputs and attributes; result tensors are allocated automatically. |
|
|
| The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model version. |
| It follows the `v1` branch as fixes land. To pin exact artifact bytes, pass a 40-character commit `revision` instead of `version`. |
|
|
| Replace each `*Data` placeholder with a typed array containing the corresponding input data. |
|
|
| ```js |
| import { getKernel } from "@huggingface/kernels"; |
| |
| const kernel = await getKernel("webgpu-kernels/ai.onnx.SpaceToDepth", { version: 1 }); |
| const { output } = await kernel({ input: { data: inputData, shape: [1, 1, 2, 4] } }, { |
| attrs: { blocksize: 2 }, |
| }); |
| ``` |
|
|
