| --- |
| library_name: kernels |
| license: apache-2.0 |
| tags: |
| - kernel |
| - webgpu |
| - wgsl |
| --- |
| # com.microsoft.GatedAdd |
|
|
| `com.microsoft` · ONNX Runtime contrib operator · contrib since_version 1 |
| |
| ## Description |
| |
| Adds `Y`, scaled by a per-row `gate`, to `X`: `output = X + round_to_T(Y * gate)`. `X` and `Y` have shape `(..., C)`; `gate` has the same rank with a trailing dimension of 1, so one value covers each row of `C` channels. Rounding the product to `T` before the addition preserves the semantics of a separate `Mul` followed by `Add`. Bfloat16 is not implemented. |
| |
| See the [ONNX Runtime `GatedAdd` contrib-operator spec](https://github.com/microsoft/onnxruntime/blob/main/docs/ContribOperators.md#com.microsoft.GatedAdd) for the reference semantics. |
| |
| ## Inputs |
| |
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `X` | `X` | `T` | — | — | Unscaled input with shape `(..., C)`. Any rank of at least 1 is accepted; only the trailing channel axis is distinguished. | required | |
| | `Y` | `Y` | `T` | — | — | Input scaled by the gate, with the same shape as `X`. | required | |
| | `gate` | `gate` | `T` | — | — | Per-row gate with shape `(..., 1)`: the same rank and leading dimensions as `X`, with a trailing dimension of 1 that broadcasts over the `C` channels. | required | |
| |
| ## Outputs |
| |
| | Name | Bind key | Logical dtype | Rank | Shape | Description | Presence | |
| | --- | --- | --- | --- | --- | --- | --- | |
| | `output` | `output` | `T` | same as `X` | same as `X` | Gated sum `X + round_to_T(Y * gate)`, with the same shape as `X`. | required | |
| |
| ## Type constraints |
| |
| | Variable | Allowed dtypes | |
| | --- | --- | |
| | `T` | `float32`, `float16` | |
| |
| ## Files |
| |
| - [`metadata.json`](build/webgpu/metadata.json) — kernel metadata (id, digests, 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 |
| - [`gated-add.wgsl.jinja`](build/webgpu/gated-add.wgsl.jinja) |
| |
| ## Use with `@huggingface/kernels` |
| |
| The loader derives every required output's shape and logical dtype from the manifest contract and this call. |
| It then allocates the result tensors automatically. |
| |
| The `version: 1` option selects the published kernel contract; it is independent of any operator opset, contrib `since_version`, or model 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/com.microsoft.GatedAdd", { version: 1 }); |
| const { output } = await kernel({ |
| X: { data: XData, shape: [2, 3] }, |
| Y: { data: YData, shape: [2, 3] }, |
| gate: { data: gateData, shape: [2, 1] }, |
| }); |
| ``` |
|
|
