scatter_op

→TensorImplPtr

int scatter_op(const int & base, int dim, const int & indices, const int & src)

source edit

Python wrappers

pylucid.ops.scatterfunction

Overwrite-semantics scatter: $\mathrm{out}[..., \mathrm{idx}, ...] = \mathrm{src}$ .

Composite over gather_op + sub_op + scatter_add_op. The overwrite is encoded as a scatter-add of the delta $\mathrm{src} - \mathrm{base}[\mathrm{idx}]$ , so the result is identical to assignment but reuses the existing additive scatter kernel. Gradient flows through ScatterAddBackward, SubBackward, and GatherBackward.

Math

y_i = \begin{cases} \mathrm{src}_j & \text{if } i = \mathrm{idx}_j \text{ for some } j \\ \mathrm{base}_i & \text{otherwise} \end{cases}

See Also

scatter_add_op — additive variant; canonical primitive.

Parameters

baseTensorImplPtr

Tensor to scatter into. Provides the output shape and dtype.

dimint

Axis along which indices selects.

indicesTensorImplPtr

Integer tensor (int32 or int64) with the same shape as src.

srcTensorImplPtr

Source values written into base at the positions given by indices.

Returns

TensorImplPtr

A new tensor with the same shape and dtype as base.

Raises

Failure

If any input is null or indices is not an integer dtype.

IndexError

If dim is out of range.

Notes

Duplicate indices yield undefined ordering — the last write wins guarantee of true scatter is not preserved here because the underlying primitive is additive. Use disjoint indices for deterministic behaviour.