RawBayerToBgraOp

RawBayerToBgraOp converts an RG10 Bayer frame into a display-ready CUDA BGRA image. It combines optical-black subtraction, per-frame gray-world white balance, Bayer demosaicing, display gain and gamma encoding in one CUDA processing stage.

Bayer processing

For help adapting Bayer processing to your sensor and connecting the output to a display, contact RidgeRun.

Input arguments

Argument	Default and meaning
`width`, `height`	Positive RG10 input dimensions.
`optical_black`	`100`; black-level value subtracted after raw decoding.
`bayer_grid`	`1`; RGGB, the pattern used by the IMX274 RG10 path. Values 0–3 select the supported Bayer layouts.
`cuda_device`	CUDA device used by the kernels; defaults to `0`.
`raw_shift`	`0`; right shift applied before the 10-bit mask.
`display_gain`, `gamma`	Both default to `1.0`; display scaling and gamma encoding applied to the output.

Ports and data contract

Port	Memory	Queue	Trigger	Contract
`input`	CUDA-visible memory, contiguous	SDK default	`OnEach`	U16 rank-3 tensor with shape `[height, width, 1]`. The valid raw sample is the lower 10-bit value after `raw_shift`.
`output`	CUDA device, runtime pool	N/A	Same activation	New U8 rank-3 tensor with shape `[height, width, 4]` in BGRA order.

Processing behavior

The operator first decodes and clamps the RG10 values after optical-black subtraction. It accumulates red, green and blue channel averages, calculates bounded gray-world gains, then performs local Bayer interpolation. The final kernel applies display gain and gamma and writes BGRA bytes. The output is not resized and retains the input dimensions.

V4L2UserPtrSourceOp -> RawBayerToBgraOp -> X11DisplayOp
V4L2UserPtrSourceOp -> RawBayerToBgraOp -> DrmBgraDisplayOp

The operator emits on an input-driven contract and forwards frame metadata. It is the fused Holoscan 5 replacement for the separate historical 4.5 raw-processing, demosaic and gain stages; those application-local names are not registered operators in the current RidgeRun repository.

Build

Follow Holoscan RidgeRun Build and Installation for the module build and installation. The operator-specific include, target, and runtime requirements are listed below.

Include raw_bayer_to_bgra/raw_bayer_to_bgra.hpp and link holoscan::raw_bayer_to_bgra.

Example

The capture source and converter must use the same dimensions and CUDA device. The repository's display applications follow this pattern:

auto camera = graph.op<V4L2UserPtrSourceOp>(
    "camera", "/dev/video0", width, height, 60, 0);
auto convert = graph.op<RawBayerToBgraOp>(
    "convert", width, height, 100, 1, 0, 0, 1.0F, 1.0F);
graph.add_flow(camera->frame, convert->input);

compile_options.deployment.bind_tensor_output_device(
    holoscan::TensorOutputDevicePlacement{
        .operator_path = "convert",
        .output_port = "output",
        .device = holoscan::DeviceId{0}});

The bayer_grid values are 0=BGGR, 1=RGGB, 2=GRBG and 3=GBRG. raw_shift is applied before extracting the ten-bit value. The corrected value is clamped after subtracting optical_black; the display scale uses the remaining range 1023 - optical_black.

Raw format assumptions

The source must provide contiguous CUDA-visible RG10 data with the configured Bayer phase. This operator does not unpack MIPI RAW10 byte packing or repair an incorrect sensor mode.

Source reference

You can find the source code at RawBayerToBgraOp.

Related pages

V4L2UserPtrSourceOp supplies the zero-copy-oriented raw input.
RawBayerToYoloTensorOp is the inference-oriented fused raw preprocessing path.

❯