AutoTVM or Quantization for Raspberry Pi fails

Hi, I already opened an issue on GitHub (Issue), but now I am not entirely sure, if it is an issue or a user error.

I want to create auto-tuned and quantized standalone bundles for the Raspberry Pi 4. I was able to cross-compile ONNX and H5 files into binaries following the bundle example in /apps.

But when I compile a auto-tuned or quantized network, it cannot be executed on the device and just outputs:

function handle for fused_X not found

where the X usually equals terms like:

  • transpose_multiply_round_clip_cast_layout_transform
  • nn_contrib_conv2d_NCHWc_add_right_shift_clip_cast_1
  • cast_multiply_add_nn_relu_1
  • nn_contrib_conv2d_NCHWc_add_right_shift_clip_cast
  • nn_max_pool2d_multiply_round_clip_cast
  • cast_multiply_add_nn_relu

What can I do to fix that?