ai.onnx.ScatterND

ai.onnx · standard ONNX operator · ONNX opset ≥ 18

Description

Copies data and then scatters values from updates into the copy at positions specified by indices, producing an output with the same shape as data. Each entry in indices is a k-tuple that addresses either a single element (when k equals the rank of data) or a slice; an optional reduction attribute replaces the default overwrite with an elementwise add, mul, max, or min accumulation.

See the ONNX ScatterND spec for the reference semantics.

Inputs

Name	Logical dtype	WebGPU storage	Rank	Shape	Description	Presence
`data`	`T`	runtime-selected; narrow integers and bool use 32-bit slots	—	—	Source tensor of rank r >= 1 that is copied to form the base of the output.	required
`indices`	`I`	`int32`	—	—	Logical int64 index tensor of rank q >= 1 whose last dimension gives signed k-tuples that address into `data`; WebGPU stores it as int32.	required
`updates`	`T`	runtime-selected; narrow integers and bool use 32-bit slots	—	—	Values to scatter into the output, of rank q + r - k - 1, matching the leading (q-1) index dimensions and the trailing (r-k) data slice dimensions.	required

Outputs

Name	Logical dtype	Rank	Shape	Description	Presence
`output`	`T`	same as `data`	same as `data`	Result tensor of rank r with the same shape as `data`, with selected positions overwritten or reduced by `updates`.	required

Attributes

Default values (overridable per request):

Attribute	Default	Description
`reduction`	`"none"`	Reduction to apply when writing each update value: `none` (overwrite, no duplicate index tuples), `add`, `mul`, `max`, or `min`.

Type constraints

Variable	Allowed dtypes
`T`	`float32`, `float16`, `int32`, `uint32`, `int8`, `uint8`, `int16`, `bool`
`I`	`int64`

Files

metadata.json — kernel metadata (id, digests, per-variant templates, provenance)
manifest.json — the op contract (source of truth)
test.json — correctness cases
bench.json — benchmark + tuning cases
datamove-scatter-row.wgsl.jinja
scatter-f16-f32-convert.wgsl.jinja
scatter-flat-copy.wgsl.jinja
scatter-narrow-wrap.wgsl.jinja
scatter-nd-reduction-atomic.wgsl.jinja
scatter-nd-reduction-f16-packed-atomic.wgsl.jinja
scatter-nd-reduction-slice.wgsl.jinja
scatter-nd.wgsl.jinja
scatter-scalar-add.wgsl.jinja

Use with `@huggingface/kernels`

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.

import { getKernel } from "@huggingface/kernels";

const kernel = await getKernel("webgpu-kernels/ai.onnx.ScatterND", { version: 1 });
const { output } = await kernel({
  data: { data: dataData, shape: [2, 2] },
  indices: { data: indicesData, shape: [2, 1] },
  updates: { data: updatesData, shape: [2, 2] },
});

Downloads last month: -

kernel

webgpu

wgsl

apache-2.0

WebGPU

Requires WebGPU support. See the compatibility table.