Profiler Output JSON Structure - Profiler Output JSON Structure - 6.3 English - XD056

The profiler captures timing data for the preprocess, inference, and postprocess stages of the pipeline. When the pipeline concludes, the profiling results are written to a JSON file with the following structure:

General Metrics:

• backend: The inference backend runtime used (for example, "onnxruntime:vitisai-ep", "vart").

• duration_sec: Total duration of the pipeline execution in seconds.

• frames: Total number of frames processed.

• fps_avg: Average frames per second (FPS) achieved during execution.

Per-Stage Metrics:

For each stage (preprocess, inference, postprocess), the following timing information is captured:

• init_us: Initialization time for the stage in microseconds.

• deinit_us: De-initialization time for the stage in microseconds.

• avg_us: Average execution time per frame in microseconds.

• min_us: Minimum execution time observed in microseconds.

• max_us: Maximum execution time observed in microseconds.

• samples: Number of samples (frames) processed in this stage.

Sample Profiler Output:

{
  "backend": "onnxruntime:vitisai-ep",
  "duration_sec": "8.250",
  "frames": 1062,
  "fps_avg": "128.721",
  "preprocess": {
    "init_us": 26431,
    "deinit_us": 239,
    "avg_us": 350,
    "min_us": 345,
    "max_us": 697,
    "samples": 1066
  },
  "inference": {
    "init_us": 1596981,
    "deinit_us": 2,
    "avg_us": 7641,
    "min_us": 7597,
    "max_us": 8140,
    "samples": 1063
  },
  "postprocess": {
    "init_us": 79,
    "deinit_us": 147,
    "avg_us": 74,
    "min_us": 67,
    "max_us": 146,
    "samples": 1062
  }
}

This profiling data helps identify performance bottlenecks across different stages of the inference pipeline, enabling optimization and performance tuning.