| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.Col2Im |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 18 |
|
|
| ## Description |
|
|
| Rearranges column blocks back into a batched multidimensional image. Takes a 3-D input of shape `[N, C * product(block_shape), L]` (where `L` is the number of blocks) and accumulates overlapping block contributions into the output image using the specified `block_shape`, `strides`, `pads`, and `dilations`. |
|
|
| See the [ONNX `Col2Im` spec](https://onnx.ai/onnx/operators/onnx__Col2Im.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Logical dtype | WebGPU storage | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `input` | `T` | same as logical dtype | `3` | — | Column-block data tensor of shape `[N, C * product(block_shape), L]` to be folded back into an image. | required | |
| | `image_shape` | `I` | `uint32` | `1` | — | Logical int64 metadata tensor specifying the output spatial dimensions (for example, `[H, W]` for 2-D); non-negative values use uint32 WebGPU storage. | required | |
| | `block_shape` | `I` | `uint32` | `1` | — | Logical int64 metadata tensor specifying the positive block size on each spatial axis (for example, `[H_block, W_block]` for 2-D); values use uint32 WebGPU storage. | required | |
|
|
| ## Outputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `output` | `T` | derived | — | Output image tensor produced by accumulating rearranged column blocks. Its rank is two greater than the length of `image_shape`. | required | |
|
|
| ## Attributes |
|
|
| Attributes and default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `dilations` | — | Dilation factors for each spatial axis; defaults to one on every axis. | |
| | `pads` | — | Padding at the beginning of every spatial axis followed by padding at the end of every spatial axis; defaults to zeros. | |
| | `strides` | — | Stride factors for each spatial axis; defaults to one on every axis. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32` | |
| | `I` | `int64` | |
|
|
| ## 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. |
|
|
| - `identity_copy_vec4` — A 1x1 block with unit strides and dilations and no padding folds every column straight onto its pixel, so the column tensor and the image are the same bytes: one flat vec4 copy, elided outright when the output is a view of the input. The block product is read off the shapes (column channels equal image channels only when every block extent is one). |
| - `identity_copy` — Copies 1x1 blocks with scalar loads and stores when the element count is not a multiple of four. |
|
|
| ## 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 |
| - [`col2im-nd.wgsl.jinja`](build/webgpu/col2im-nd.wgsl.jinja) |
| - [`datamove-elementwise-copy.wgsl.jinja`](build/webgpu/datamove-elementwise-copy.wgsl.jinja) |
| - [`datamove-flat-copy.wgsl.jinja`](build/webgpu/datamove-flat-copy.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.3 |
| ``` |
|
|
| Outputs with inferable metadata are allocated automatically. Explicit `outputs` entries request optional results or provide metadata that cannot be inferred from the supplied inputs and attributes. |
|
|
| This example supplies explicit metadata for: |
|
|
| - `output` |
|
|
| 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.Col2Im", { version: 1 }); |
| // Explicit destinations request optional results or supply metadata that cannot be inferred. |
| const { output } = await kernel({ |
| input: { data: inputData, shape: [1, 9, 1] }, |
| image_shape: { data: image_shapeData, shape: [2] }, |
| block_shape: { data: block_shapeData, shape: [2] }, |
| }, { |
| outputs: { output: { shape: [1, 1, 3, 3], dtype: "float32" } }, |
| }); |
| ``` |
|
|
