| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.Gather |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 13 |
|
|
| ## Description |
|
|
| Gathers entries from `data` along the specified `axis` using an `indices` tensor, producing an output of rank `q + (r - 1)` where `r` is the rank of `data` and `q` is the rank of `indices`. Each index value selects a `(r-1)`-dimensional slice from `data` along `axis`, and the resulting slices are arranged into the output tensor with the `q` index dimensions replacing the gathered axis. |
|
|
| See the [ONNX `Gather` spec](https://onnx.ai/onnx/operators/onnx__Gather.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `data` | `T` | — | — | Tensor of rank `r >= 1` to gather entries from. | required | |
| | `indices` | `I` | — | — | Integer index tensor of any rank q; each value must be in `[-s, s-1]` where `s` is the size of `data` along `axis`. | required | |
|
|
| ## Outputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `output` | `T` | derived | derived | Output tensor of rank q + (r - 1) containing the gathered slices. | required | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | `0` | Axis of `data` along which to gather; negative values count from the back. Accepted range is `[-r, r-1]` where `r = rank(data)`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16`, `uint32`, `int32`, `int16`, `uint8`, `int8`, `bool`, `int64` | |
| | `I` | `int32` | |
|
|
| ## 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 |
| - [`datamove-gather-row.wgsl.jinja`](build/webgpu/datamove-gather-row.wgsl.jinja) |
| - [`gather.wgsl.jinja`](build/webgpu/gather.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.Gather", { version: 1 }); |
| const { output } = await kernel({ |
| data: { data: dataData, shape: [4] }, |
| indices: { data: indicesData, shape: [4] }, |
| }); |
| ``` |
|
|
