qnn-context-binary-generator

soc id and hexagon arch id

See

• https://github.com/alibaba/MNN/wiki/npu

• https://github.com/tianxf99/executorch/blob/main/backends/qualcomm/serialization/qc_compiler_spec.fbs

• https://github.com/tianxf99/executorch/blob/main/backends/qualcomm/utils/utils.py#L963

• https://docs.qualcomm.com/bundle/publicresource/topics/80-63442-10/QNN_general_overview.html

• SA: Snapdragon Automotive

• SM: Snapdragon Mobile

Hardware	SOC ID	Hexagon arch
865 (SM8250)	21	66
SA8295	39	68
888 (SM8350)	28	68
8 Gen 1 (SM8450)	36	69
8+ Gen 1 (SM8475)	42	69
8 Gen 2 (SM8550)	43	73
8 Gen 3 (SM8650)	57	75
7 Gen 2	54	75
8 Elite (SM8750)	69	79
8 Gen 5	79	80
8 Elite Gen 5 SM8850	87	81
SSG2115P	46	73

config file

See https://github.com/alibaba/MNN/blob/6686bed19bf8a1cd61878a82e53d20c482fffdb3/tools/cpp/MNN2QNNModel.cpp#L350

It generates two json files:

context_config.json

{
  "backend_extensions": {
    "shared_library_path": "/content/qairt/2.33.0.250327/lib/x86_64-linux-clang/libQnnHtpNetRunExtensions.so",
    "config_file_path": "./htp_backend_extensions.json"
  }
}

htp_backend_extensions.json

{
  "graphs": [
    {
      "vtcm_mb": 8,
      "graph_names": ["model_op17_static_93_quantized"],
      "O": 3.0,
      "fp16_relaxed_precision": 1,
      "weights_packing": true,
      "hvx_threads": 4
    }
  ],
  "devices": [
    {
      "soc_id": 79,
      "dsp_arch": "v80",
      "cores": [
        {
          "core_id": 0,
          "perf_profile": "burst",
          "rpc_control_latency": 100
        }
      ]
    }
  ],
  "context": {
    "weight_sharing_enabled": true
  }
}

Run command

"${QNN_SDK_ROOT}/bin/x86_64-linux-clang/qnn-context-binary-generator" \
  --model=./libmodel.so \
  --backend=${QNN_SDK_ROOT}/lib/x86_64-linux-clang/libQnnHtp.so \
  --binary_file=./model.bin \
  --config_file=./context_config.json