FakesinkOp

FakesinkOp consumes Tensor samples without copying or inspecting their payload. It is the terminal sink used when a benchmark must measure capture, preprocessing or inference without adding display work.

Pipeline profiling

Contact RidgeRun for help finding which stages use the most processing time and improving your pipeline's performance.

Input arguments

The optional verbose constructor argument enables logging of received tensors. The sink has no output port and does not allocate a replacement Tensor.

Ports and data contract

Port	Memory	Queue	Trigger	Contract
`input`	Tensor memory supplied by the upstream operator	1	`OnEach`	Receives each selected sample and releases it after recording the event. No Holoscan output.

V4L2UserPtrSourceOp -> RawBayerToYoloTensorOp -> TensorRtInferenceOp -> FakesinkOp

The same sink can terminate other Tensor graphs, including the SDK-owned InferenceOp path.

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 fakesink_op/fakesink_op.hpp and link holoscan::fakesink_op.

Example

The received Sample<Tensor> is local to compute(). When it goes out of scope, the sink releases the plan-owned Tensor storage. No synchronization, host copy, Tensor decode or inference postprocessing occurs. With verbose=true, the implementation prints the buffer count, shape, dtype, device, byte count, sequence and frame ID; this is useful for checking a graph but adds logging overhead.

The sink also records the receive and release interval in the shared RidgeRun flow trace under the sink stage. That trace event is independent of Glass-to-Glass latency because no display or optical measurement is performed.

auto sink = graph.op<holoscan::ridgerun::FakesinkOp>("fakesink", false);
graph.add_flow(inference->output, sink->input);

Not a display sink

FakesinkOp does not render, decode detections, run NMS or transform the inference output. Its timing represents a software endpoint, not Glass-to-Glass latency.

Source reference

You can find the source code at FakesinkOp.

Related pages

TensorRtInferenceOp documents the inference stage before the sink.
Holoscan vs GStreamer defines the headless benchmark interpretation.

❯