The complete pipeline FPS is measured with a consistent source and sink.
GStreamer tracing or debug logs show that the bottleneck is not file I/O, decode, overlay, or the sink.
The
vvas_xinferprofiler is enabled and the JSON output has been reviewed.The profiler
backendmatches the expected runtime.Accelerated inference uses either direct VART or ONNX Runtime with the Vitis AI EP; ONNX Runtime CPU EP is used only when software fallback is intended.
Batch behavior and
batch-timeoutare appropriate for the latency and throughput target.postprocesstime is not dominating thevvas_xinferprofiler output.The pipeline uses the intended zero-copy or DMA-BUF path where applicable.