| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # ai.onnx.ScatterElements |
|
|
| `ai.onnx` · standard ONNX operator · ONNX opset ≥ 18 |
|
|
| ## Description |
|
|
| Produces a copy of `data` with values updated at positions given by `indices` along the specified `axis`. For each entry in `updates`, the axis coordinate comes from `indices` while all other coordinates come from the entry's own position in `updates`. An optional `reduction` (`add`, `mul`, `max`, `min`) combines updates with existing values instead of overwriting; with `none`, duplicate indices are not allowed. |
|
|
| See the [ONNX `ScatterElements` spec](https://onnx.ai/onnx/operators/onnx__ScatterElements.html) for the reference semantics. |
|
|
| ## Inputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `data` | `T` | — | — | Input tensor of rank r >= 1 that is copied to form the output base. | required | |
| | `indices` | `I` | — | — | Integer index tensor of the same rank as `data`; each value selects a position along `axis`. | required | |
| | `updates` | `T` | — | — | Values to scatter, same rank and shape as `indices`. | required | |
|
|
| ## Outputs |
|
|
| | Name | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | |
| | `output` | `T` | same as `data` | same as `data` | Copy of `data` with scattered updates applied; same shape as `data`. | required | |
|
|
| ## Attributes |
|
|
| Default values (overridable per request): |
|
|
| | Attribute | Default | Description | |
| | --- | --- | --- | |
| | `axis` | `0` | Which axis to scatter on; negative values count from the back. Accepted range is `[-r, r-1]` where `r = rank(data)`. | |
| | `reduction` | `"none"` | Reduction to apply when writing updates: `none` (overwrite, no duplicate indices), `add`, `mul`, `max`, or `min`. | |
|
|
| ## Type constraints |
|
|
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16`, `int32`, `uint32`, `int8`, `uint8`, `int16`, `bool` | |
| | `I` | `int32` | |
|
|
| ## Device requirements |
|
|
| Some implementation variants require `subgroups`. These are route-specific capabilities, not package-wide requirements; availability also depends on the request shape and dtype. |
|
|
| ## 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 + tuning cases |
| - [`scatter-elements-f32-add-axis0-histogram.wgsl.jinja`](build/webgpu/scatter-elements-f32-add-axis0-histogram.wgsl.jinja) |
| - [`scatter-elements-reduction-atomic.wgsl.jinja`](build/webgpu/scatter-elements-reduction-atomic.wgsl.jinja) |
| - [`scatter-elements-reduction-slab.wgsl.jinja`](build/webgpu/scatter-elements-reduction-slab.wgsl.jinja) |
| - [`scatter-elements-reduction.wgsl.jinja`](build/webgpu/scatter-elements-reduction.wgsl.jinja) |
| - [`scatter-elements.wgsl.jinja`](build/webgpu/scatter-elements.wgsl.jinja) |
| - [`scatter-f16-f32-convert.wgsl.jinja`](build/webgpu/scatter-f16-f32-convert.wgsl.jinja) |
| - [`scatter-flat-copy.wgsl.jinja`](build/webgpu/scatter-flat-copy.wgsl.jinja) |
| - [`scatter-narrow-wrap.wgsl.jinja`](build/webgpu/scatter-narrow-wrap.wgsl.jinja) |
|
|
| ## Use with `@huggingface/kernels` |
|
|
| ```sh |
| npm install --save-exact @huggingface/kernels@0.0.1-preview.2 |
| ``` |
|
|
| 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.ScatterElements", { version: 1 }); |
| const { output } = await kernel({ |
| data: { data: dataData, shape: [3] }, |
| indices: { data: indicesData, shape: [2] }, |
| updates: { data: updatesData, shape: [2] }, |
| }); |
| ``` |
|
|
