Is this page helpful?

Slice#

The slice operation extracts a strided window of a tensor along each dimension:

\( Y = X[\text{start}_0:\text{limit}_0:\text{step}_0,\ \ldots] \)

where each dimension uses a half-open interval [start, limit) and an optional slice stride (step) per axis. The output length along axis i is ceil((limit_i - start_i) / step_i) when step_i > 0.

Backend operation (cuDNN 9.22+)#

On cuDNN 9.22.0 and newer, the node can lower to the native backend slice operation and materializes both input and output tensor descriptors. On older versions, the frontend may use a pointer-offset style lowering for compatible cases.

C++ API#

std::shared_ptr<Tensor_attributes>
Slice(std::shared_ptr<Tensor_attributes> input, Slice_attributes);

Slice_attributes setters:

Slice_attributes&
set_slices(std::vector<std::pair<int64_t, int64_t>> const value);  // [start, limit) per dim

Slice_attributes&
set_strides(std::vector<int64_t> const value);  // step per dim; must be > 0; defaults to 1

Slice_attributes&
set_name(std::string const&);

Slice_attributes&
set_compute_data_type(DataType_t value);

If set_strides is omitted, every step defaults to 1. If provided, the vector must cover each sliced dimension consistently with set_slices.

Python API#

  • slice(input, slices=[], compute_data_type=cudnn.data_type.NOT_SET, name="")

    • input: tensor to slice.

    • slices: list of slice objects, one per dimension. Per-axis step is taken from each slice’s step (default 1), after normalization for that axis length (same idea as sequence[sl] in Python).

Example:

input_tensor = graph.tensor(dim=[4, 8, 16], stride=[128, 16, 1], data_type=cudnn.data_type.float)
sliced_tensor = graph.slice(
    input_tensor,
    slices=[slice(1, 3), slice(2, 6, 2), slice(0, 16)],
    name="my_slice",
    compute_data_type=cudnn.data_type.float,
)