ReformatOp

ReformatOp converts CUDA image storage between interleaved HWC/NHWC and planar CHW/NCHW layouts. It produces a new contiguous tensor with the same image dimensions, channel count, scalar type, and capture metadata.

Layout describes where the channel values are stored. A BGR image remains BGR; changing the layout does not swap channels, normalize values, or convert the numeric type.

Input arguments

ReformatOpParams has one field, mode, defaulting to kInterleavedToPlanar. Its ReformatMode choices are:

Mode	Shape Transformation
`kInterleavedToPlanar`	`[H,W,C] → [C,H,W]`; `[N,H,W,C] → [N,C,H,W]`.
`kPlanarToInterleaved`	`[C,H,W] → [H,W,C]`; `[N,C,H,W] → [N,H,W,C]`.

The field is CompileTime, frozen for one accepted plan. Configure it before compilation and compile a new plan to change direction. It is not marked as an admitted runtime-mutable field. Output capacity follows the connected input's finite byte bound. The operator has no separate image-capacity constructor.

Ports and data contract

Port	Memory	Queue	Trigger	Contract
`input`	CUDA device	1	`OnEach`	Rank-three or rank-four image using a supported scalar type and the configured interleaved or planar layout.
`output`	CUDA device	N/A	Same activation	Same image data in the requested layout. Capture metadata is preserved.

The layout and direction are configured explicitly. Interleaved channels and planar pixels must be packed as required by the selected layout.

Example

With graph already declared and <holoscan/vision/reformat_op.hpp> included:

auto reformat = graph.op<holoscan::vision::ReformatOp>("to-planar");
reformat.set_params(holoscan::vision::ReformatOpParams{
    .mode = holoscan::vision::ReformatMode::kInterleavedToPlanar,
});

An HWC [2,2,3] result becomes CHW [3,2,2]. The first output plane contains the first channel's four pixels, then the second and third planes follow. Connect the interleaved CUDA result to reformat->input; its output is suitable only for a consumer that accepts the resulting planar representation.

For the reverse direction, set kPlanarToInterleaved and supply CHW/NCHW input. Reformat changes storage order through GPU work; it is not a metadata-only view.

Build

Follow Build the SDK on AGX Orin for the SDK build and installation. Link the installed holoscan::vision target in the application.

Related examples and concepts

Examples

Vision Pipeline Example supplies a complete checked source, transformations and validating sinks.

Concepts

Dynamic Graph Behavior distinguishes authoring parameters from admitted runtime publication.
Building Graphs and Queue Policies explains buffering and pressure.
NormalizeOp applies channel arithmetic before the layout boundary.
HolovizOp describes the display-facing tensor formats.

Source You can find the source code at ReformatOp.

Official documentation

NVIDIA Holoscan SDK User Guide provides the official operator and application reference.

❯