| |
| |
| """How much of each accelerator encoder the phone accelerators can take, by ONNX Runtime's own mobile-usability |
| check (python -m onnxruntime.tools.check_onnx_model_mobile_usability): for NNAPI, Core ML NeuralNetwork and Core ML |
| MLProgram, the number of partitions, the nodes they cover and the operators they cannot run. One partition covering |
| every node means the accelerator runs the whole encoder in one call; several mean round trips through the CPU. |
| |
| The check reads the graph only. It cannot see a phone's driver, and it lists caveats it does not verify (e.g. NNAPI |
| takes only 2-D Conv: a graph with 1-D convolutions passes the check and still splits on a device). What a runtime |
| actually did is in results/android/emulator_x86_64.json (NNAPI, "placement") and results/gates/qnn_soc*.json (QNN). |
| |
| python3 gate/coverage.py [models/anuvani_lid.npu-15s.encoder.onnx ...] (default: every *.encoder.onnx) |
| Report -> results/gates/accelerator_coverage.json; the checker's full output -> results/gates/usability/<model>.txt |
| """ |
| import glob |
| import json |
| import os |
| import re |
| import subprocess |
| import sys |
|
|
| sys.path.insert(0, os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "tools")) |
| import common |
|
|
| SECTIONS = {"NNAPI": "nnapi", "CoreML NeuralNetwork": "coreml_neuralnetwork", "CoreML MLProgram": "coreml_mlprogram"} |
|
|
|
|
| def parse(text): |
| out = {} |
| for title, key in SECTIONS.items(): |
| m = re.search(rf"Checking {title}\n(.*?)(?:={{8,}}|\Z)", text, re.S) |
| if not m: |
| continue |
| sec = m.group(1) |
| p = re.search(r"(\d+) partitions with a total of (\d+)/(\d+) nodes", sec) |
| u = re.search(r"Unsupported ops: (.*)", sec) |
| ok = re.search(r"Model should perform well with .* as is: (YES|NO)", sec) |
| out[key] = {"partitions": int(p.group(1)) if p else None, |
| "nodes_supported": int(p.group(2)) if p else None, |
| "nodes_total": int(p.group(3)) if p else None, |
| "unsupported_ops": [o.strip() for o in u.group(1).split(",")] if u else [], |
| "checker_verdict_perform_well": ok.group(1) if ok else None} |
| return out |
|
|
|
|
| def main(): |
| models = sys.argv[1:] or sorted(glob.glob(os.path.join(common.REPO, "models", "*.encoder.onnx"))) |
| import onnxruntime as ort |
| rep = {"_doc": __doc__.strip().split("\n\n")[0], "onnxruntime": ort.__version__, "models": {}} |
| os.makedirs(os.path.join(common.REPO, "results", "gates", "usability"), exist_ok=True) |
| for m in models: |
| r = subprocess.run([sys.executable, "-m", "onnxruntime.tools.check_onnx_model_mobile_usability", m], |
| capture_output=True, text=True) |
| text = r.stdout + r.stderr |
| text = re.sub(r"^INFO:\s+", "", text, flags=re.M) |
| name = os.path.basename(m) |
| with open(os.path.join(common.REPO, "results", "gates", "usability", name[:-5] + ".txt"), "w") as f: |
| f.write(text) |
| rep["models"][name] = {"sha256": common.sha256(m), **parse(text)} |
| print(name, json.dumps(rep["models"][name])) |
| out = os.path.join(common.REPO, "results", "gates", "accelerator_coverage.json") |
| json.dump(rep, open(out, "w"), indent=1) |
| print("->", os.path.relpath(out, common.REPO)) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
