This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

AM62A7-Q1: edgeai-tidl-tools ver 9.2 : Segmentation fault occurred

Part Number: AM62A7-Q1

Tool/software:

I cloned edgeai-tidl-tools ver 9.2 and completed the environment installation using source ./setup.sh. 
And we converted edgeai-yolov5's m6 model with the "python3 onnxrt_ep.py -c" command.
I tried object detection by copying the model to the TI AM62A7-Q1 board, but a segmentation fault occurred. The results of debugging with gdb are as follows. Segmentation fault. 0x0000fffff70ce640 in OrtApis::GetTensorTypeAndShape(OrtValue const*, OrtTensorTypeAndShapeInfo**) () from /usr/lib/python3.10/site-packages/onnxruntime/capi/onnxruntime_pybind11_state.so The version of onnxruntime on the board and the version of onnxruntime on the PC whose model was converted matched. I would like to inquire about a solution.
  • Hello,

    I will require more information to suggest a solution. I'm not sure if the seg-fault you are seeing happened before or after TIDL ran on the C7xMMA, but I see that it's an ONNXRuntime API that shows this issue.

    Could you please include the compile logs and the inference logs? Please use the following configuration for more logging info

    • Set debug_level=2 in the set of delegate options passed to the onnxruntime. This nominally set in common_utils.py. Use this for PC compilation and target inference
    • export TIDL_RT_DEBUG=1 # on target inference and host compilation
    • /opt/vx_app_arm_remote_log.out & #only on target inference

    I am curious if you also run into this issue when running the model with C7x emulation on the host PC.

    Note that the onnxruntime version that we use is a fork of the public onnxruntime. If you use the `pip freeze` command to check python versions, our version of onnxruntime will show as "onnxruntime-tidl". Please ensure this is the case on host and target.

    BR,
    Reese