Frontend APIOperations

Reshape

View as Markdown

The reshape operation connects an input tensor to an output tensor with a different logical shape and stride, without changing the underlying element type. The output rank and layout are usually set on the output tensor (or via Reshape_attributes::set_dim / set_stride); the node wires the reshape in the graph.

Reshape modes (ReshapeMode_t)

ModeMeaning
VIEW_ONLYDefault. View-style reshape when the layout is compatible with the backend.
LOGICALLexicographic logical reshape: use when the mapping is expressed in logical (row-major lexicographic) terms as required by the backend. Prefer this when a plain view is not sufficient for the target shape.

Mode is ignored on older cuDNN builds that do not expose reshape mode; on cuDNN 9.22+, it is passed to the backend reshape operation.

C++ API

std::shared_ptr<Tensor_attributes>
reshape(std::shared_ptr<Tensor_attributes> input, Reshape_attributes attributes);

Relevant Reshape_attributes setters:

Reshape_attributes&
set_dim(std::vector<int64_t> const& value);
Reshape_attributes&
set_stride(std::vector<int64_t> const& value);
Reshape_attributes&
set_reshape_mode(ReshapeMode_t const& value);  // VIEW_ONLY or LOGICAL
Reshape_attributes&
set_name(std::string const& value);

Python API

  • pygraph.reshape(input, name="", reshape_mode=cudnn.reshape_mode.VIEW_ONLY)
    • reshape_mode: cudnn.reshape_mode.VIEW_ONLY or cudnn.reshape_mode.LOGICAL.

Example (C++)

using cudnn_frontend::ReshapeMode_t;
auto r = fe::graph::Reshape_attributes()
             .set_name("logical_flatten")
             .set_reshape_mode(ReshapeMode_t::LOGICAL)
             .set_dim({24})
             .set_stride({1});
auto Y = graph.reshape(X, r);