VGF neural accelerator statistics in ETDump

The VGF backend emits neural accelerator profiling data through ETDump delegate metadata when runtime profiling is enabled and a supported VK_ARM_data_graph driver is available.

By default, VGF runtime profiling emits timing events only. Neural accelerator statistics collection is opt-in because the payload can include binary blobs and may increase ETDump size and profiling overhead, especially across repeated inference runs.

To enable VGF neural accelerator statistics collection, set:

EXECUTORCH_VGF_ENABLE_NEURAL_STATISTICS=1

The statistics mode defaults to mode 1. Select mode 0 or 1 explicitly with:

EXECUTORCH_VGF_NEURAL_STATISTICS_MODE=0
# or
EXECUTORCH_VGF_NEURAL_STATISTICS_MODE=1

The public values 0 and 1 map to Vulkan VK_NEURAL_ACCELERATOR_STATISTICS_MODE_STATISTICS0_ARM and VK_NEURAL_ACCELERATOR_STATISTICS_MODE_STATISTICS1_ARM, respectively. Invalid mode values produce a warning and fall back to mode 1. Set these environment variables before the VGF backend initializes its Vulkan device.

When statistics are requested, the backend enables VK_ARM_data_graph_neural_accelerator_statistics only if both the device extension and dataGraphNeuralAcceleratorStatistics feature are supported. If support is missing, inference still proceeds and the emitted metadata reports statistics as unavailable.

When this option is not set, the backend does not emit the VGF_NEURAL_STATISTICS delegate metadata event, even if general ETDump/runtime profiling is enabled.

The emitted delegate profiling event name is:

VGF_NEURAL_STATISTICS

The delegate metadata payload is a UTF-8 JSON wrapper. The wrapper schema is:

{
  "schema": "executorch.vgf.neural_statistics",
  "schema_version": 1,
  "backend": "VgfBackend",
  "api": "VK_ARM_data_graph",
  "event_name": "VGF_NEURAL_STATISTICS",
  "api_available": true,
  "data_available": true,
  "available": true,
  "reason": "",
  "segments": []
}

Each segment can contain:

{
  "segment_id": 0,
  "is_data_graph_pipeline": true,
  "statistics_bind_point_available": true,
  "statistics_memory_host_visible": true,
  "statistics_memory_host_coherent": true,
  "statistics_bind_point_reason": "",
  "debug_database": {
    "available": true,
    "is_text": false,
    "vulkan_result": 0,
    "size": 0,
    "encoding": "base64",
    "reason": "",
    "data": ""
  },
  "statistics_info": {
    "available": true,
    "is_text": true,
    "vulkan_result": 0,
    "size": 0,
    "encoding": "base64",
    "reason": "",
    "data": ""
  },
  "statistics_memory": {
    "available": true,
    "is_text": false,
    "vulkan_result": 0,
    "size": 0,
    "encoding": "base64",
    "reason": "",
    "data": ""
  }
}

We don’t parse the neural accelerator blobs. Consumers should treat debug_database, statistics_info, and statistics_memory as opaque bytes.

Reading from Inspector

from executorch.devtools.inspector import Inspector

inspector = Inspector(etdump_path="run.etdump")
records = inspector.get_vgf_neural_statistics()

for record in records:
    print(record["schema_version"], record["data_available"])
    for segment in record["segments"]:
        stats_bytes = segment["statistics_memory"]["raw_data"]
        debug_db_bytes = segment["debug_database"]["raw_data"]

If the Vulkan API, driver, or hardware support is unavailable, normal execution continues and the JSON wrapper is still emitted with:

{
  "api_available": false,
  "data_available": false,
  "available": false,
  "reason": "..."
}