scatter_add_op

→TensorImplPtr

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

source edit

Python wrappers

pylucid.ops.scatter_addfunction

User-facing scatter-add: out = base with src accumulated at positions given by indices along dim.

Math

\text{out}[i_0, \dots, i_{d-1}, \text{indices}[j], i_{d+1}, \dots] \mathrel{+}= \text{src}[i_0, \dots, j, \dots]

Parameters

baseTensorImplPtr

Tensor providing the initial values; same shape as the output.

indicesTensorImplPtr

Integer tensor indexing into base along dim. Shape matches src.

srcTensorImplPtr

Values to add.

dimint

Axis along which scattering occurs.

Returns

TensorImplPtr

Same shape and dtype as base.

Notes

Full autograd support. Gradients: $\partial / \partial \text{base} = \text{grad\_out}$ , $\partial / \partial \text{src} = \text{gather}(\text{grad\_out}, \text{dim}, \text{indices})$ .