QNN Execution Provider | Radxa Docs

[giscus[bot]]

QNN Execution Provider | Radxa Docs

https://docs.radxa.com/en/dragon/q6a/app-dev/npu-dev/qnn-onnxrt-execution-provider

[dmitry-a-l]

There is error running python3 run_qdq_model.py:

[E:onnxruntime:, qnn_execution_provider.cc:948 GetCapability] QNN SetupBackend failed qnn_backend_manager.cc:220 GetQnnInterfaceProvider Unable to load backend, error:  libQnnHtp.so: cannot open shared object file: No such file or directory

The reason is that bin/envsetup.sh incorrectly initialize LD_LIBRARY_PATH - it points to $QAIRT_SDK_ROOT/lib/x86_64-linux-clang
To run the example we need to set LD_LIBRARY_PATH=$QAIRT_SDK_ROOT/lib/aarch64-ubuntu-gcc9.4

[tangzz-radxa]

Thank you for reporting this issue with the LD_LIBRARY_PATH in bin/envsetup.sh. You are correct that for ARM64 Linux systems (like Radxa Dragon Q6A), the library path should point to $QAIRT_SDK_ROOT/lib/aarch64-ubuntu-gcc9.4 instead of $QAIRT_SDK_ROOT/lib/x86_64-linux-clang.

This appears to be an error in the QAIRT SDK setup script. We will investigate this issue and update our documentation accordingly.

In the meantime, you can manually fix this by editing the envsetup.sh file and changing the LD_LIBRARY_PATH line to point to the correct ARM64 library directory.

Thank you for bringing this to our attention!

[tangzz-radxa]

Thank you for reporting this issue. You are correct that the bin/envsetup.sh script incorrectly sets LD_LIBRARY_PATH to $QAIRT_SDK_ROOT/lib/x86_64-linux-clang when it should be $QAIRT_SDK_ROOT/lib/aarch64-ubuntu-gcc9.4 for ARM64 platforms.

This appears to be a documentation/script error that needs to be fixed. The correct path should be:

export LD_LIBRARY_PATH=$QAIRT_SDK_ROOT/lib/aarch64-ubuntu-gcc9.4:$LD_LIBRARY_PATH

I will mark this as a documentation issue that needs to be addressed in the QNN Execution Provider documentation.

[tangzz-radxa]

Thank you for reporting this issue and providing the solution! You are correct that the bin/envsetup.sh script incorrectly sets LD_LIBRARY_PATH to the x86_64 directory instead of the aarch64 directory for ARM-based systems like Radxa boards.

Your suggested fix is accurate:

export LD_LIBRARY_PATH=$QAIRT_SDK_ROOT/lib/aarch64-ubuntu-gcc9.4

This should resolve the libQnnHtp.so: cannot open shared object file error when running the QNN execution provider examples on Radxa ARM devices.

We will update the documentation to include this clarification for ARM users. Appreciate you sharing this finding!

[tangzz-radxa]

Thank you for reporting this issue. The error "libQnnHtp.so: cannot open shared object file" indicates that the Qualcomm QNN backend libraries are not properly installed or configured.

The envsetup.sh script appears to be pointing to the wrong architecture directory (x86_64-linux-clang instead of aarch64-ubuntu-gcc9.4 for ARM64 systems).

To resolve this, you need to:

1. Make sure the Qualcomm QNN SDK is properly installed on your system
2. Update the LD_LIBRARY_PATH to point to the correct architecture-specific library directory
3. Verify that the libQnnHtp.so library exists in the configured path

We will update our documentation to include clearer instructions for QNN setup on different architectures.

[tangzz-radxa]

Thanks for reporting this issue and identifying the root cause! You are correct — bin/envsetup.sh sets LD_LIBRARY_PATH to the x86_64 path, which causes libQnnHtp.so: cannot open shared object file on the aarch64 target.

We have submitted a docs fix in PR #1650 that adds the correct LD_LIBRARY_PATH before running run_qdq_model.py:

cd qairt/2.37.1.250807
source bin/envsetup.sh
export ADSP_LIBRARY_PATH=$QNN_SDK_ROOT/lib/hexagon-v${DSP_ARCH}/unsigned
export LD_LIBRARY_PATH=$QNN_SDK_ROOT/lib/aarch64-ubuntu-gcc9.4:$LD_LIBRARY_PATH

Please try with the updated instructions and let us know if it works.