How do you accurately measure end to end video latency? This toolkit provides:
- The signal generator - an app that generates a machine-readable image encoding the current wall clock time.
- The signal interpreter - an app that reads the image and determines the latency.
The generator can emit:
- A real-time updated window (for feeding input via screen capture).
The interpreter can process:
- A screen capture (for capturing a video player window).
- A stream of decoded video frames in NV12 format (stdin NV12 -> stdout JSON).
The processed video samples can include other content besides the time signal - the interpreter knows how to find the embedded signal from among other content. Just ensure that the entire signal generator window is visible in the output.
Clock synchronization between generator and interpreter is assumed. Results will be invalid if time is out of sync.
There is built-in support for automatic clock synchronization by:
- Referencing a deployed instance of dash-timeserver.
- Or specifying an NTP server with the URL
ntp://ntp.example.com.
Alternatively, running both apps on the same PC can yield acceptable results (though may be logically difficult).
There is some processing latency - perform a "dry run" by providing generator output directly to the interpreter to determine the procecssing latency on your system.
Interpreter.Gui exposes the latency in the form of Prometheus metrics, on the http://server:19891/metrics URL. You can use the metrics-port command line argument to specify a different port, allowing multiple instances of the app to publish metrics concurrently.
Requires .NET 5.0 on Windows.