Frontend APIOperations

Concatenate

View as Markdown

The Concatenate operation merges two or more tensors into one, along the specified axis. The user may also specify an in-place merge. The operation provides the capabilities of the cudnn backend’s concatenate operation.

C++ API

std::shared_ptr<Tensor_attributes>
concatenate(std::vector<std::shared_ptr<Tensor_attributes>>, Concatenate_attributes);

Concatenate attributes is a lightweight structure with inputs, outputs, and setters:

std::vector<std::shared_ptr<Tensor_attributes>> inputs;
std::unordered_map<output_names, std::shared_ptr<Tensor_attributes>> outputs;
Concatenate_attributes&
set_axis(int64_t const value)
Concatenate_attributes&
set_in_place_index(int64_t const value)

In-place concat (optional)

set_in_place_index is optional. When unset, the backend writes the concatenation result only into the output tensor. When set to index k, the concatenate operation may reuse the buffer of input k as documented for CUDNN_ATTR_OPERATION_CONCAT_INPLACE_INDEX (see the cuDNN backend concatenate descriptor). Callers must still provide a distinct output tensor in the graph API; layout and lifetime rules follow backend requirements for the chosen index.

Python API

  • pygraph.concatenate(inputs, axis, name="", in_place_index=None)
    • inputs: list of tensors to concatenate.
    • axis: dimension index to concatenate along.
    • in_place_index: optional int; omit for out-of-place behavior.

Requirements

Graph concatenate lowering requires a sufficiently new cuDNN backend (see frontend release notes for the minimum version).