Snapdragon Neural Processing Engine (SNPE) is a Qualcomm Snapdragon software accelerated runtime for the execution of deep neural networks.

Quantization Scheme

8/16 bit per-layer asymmetric linear quantization.

\[\begin{equation} q = \mathtt{clamp}\left(\left\lfloor R * \dfrac{x - cmin}{cmax - cmin} \right\rceil, lb, ub\right) \end{equation}\]

where \(R\) is the integer range after quantization, \(cmax\) and \(cmin\) are calculated range of the floating values, \(lb\) and \(ub\) are bounds of integer range. Taking 8bit as an example, R=255, [lb, ub]=[0,255].

Deploy on SNPE


  • Install SNPE SDK from QualComm (Suggest Ubuntu 18.04)


  • Convert PyTorch checkpoint to snpe_deploy.onnx and dump clip ranges to snpe_clip_ranges.json:

    1from mqbench.convert_deploy import convert_deploy
    2input_dict = {'x': [1, 3, 224, 224]}
    3convert_deploy(solver.model.module, BackendType.SNPE, input_dict)
  • Convert .onnx file to .dlc format (supported by SNPE):

    1# Note that, the `.json` file contains activation ranges for quantization, but it's required here although the model hasn't been quantized now.
    2snpe-onnx-to-dlc --input_network ./snpe_deploy.onnx --output_path ./snpe_deploy.dlc --quantization_overrides ./snpe_clip_ranges.json
  • Quantize the model with parameters overridden:

    1# The `data.txt` records paths to image data for calibration (not important since we will override parameters) which will be loaded by `numpy.fromfile(dtype=np.float32)` and have shape of `(224, 224, 3)`. And this file is required for test.
    2# Now we get the final model `snpe_deploy_quantized.dlc`
    3snpe-dlc-quantize --input_dlc ./snpe_deploy.dlc --input_list ./data.txt --override_params  --bias_bitwidth 32