How do i force the use of a specific onnxruntime.dll ?

[ekrogh]

In my windows there are lots of onnxruntime.dll.

My c++ programs insist on using C:\WINDOWS\SYSTEM32\onnxruntime.dll though i have built onnxruntime from source and added the path to onnxruntime.lib and onnxruntime.dll to the vs 2022 linker paths.

[snnn]

copy onnxruntime.dll to the same folder where your main exe lives.

[V3nt1n]

Is it possible somehow to explicitly define the path to the folder where te onnxruntime.dll is stored (Win with Cpp API)?
For example any solution of this kind:

1. Before code execution:

• Open CMD
• set PATH=D:\project\onnxruntime-win-x64-1.20.1\lib;%PATH%
• inference.exe
  2.Force the path definition within the inference.cpp code:
• SetDllDirectoryW(L"D:\project\onnxruntime-win-x64-1.20.1\lib") //explicitly define the path

(Just to mention I'm trying to create a FMU (FMI Standard https://fmi-standard.org/) to perform inference with onnx. Because this standard all the needed resources to run the FMU must be within the FMU self contained resources folder, like NN.onnx, onnxruntime.dll, data.txt, etc.)

Best regards & thnkx for your insight

[snnn]

Do not rely on "PATH" to find DLLs. For example, python.exe totally ignores it.
Copy the DLLs to the same folder as your "inference.exe". Then it should work.

[V3nt1n]

Hello, snnn. Thnx for your answer.
I'm quite new to this topic, I appreciate so much if you could provide in detail step-by-step info...

You are right If I just copy the onnxruntime.dll to the inference.exe folder, it works.
But I'm trying to create an FMU. The FMU self contain all the files(infernece.exe, onnxruntime.dll, NN.onnx, etc) . So when the FMU is launch, it creates a temp folder with all the files. It is executed within this temp folder.
I need that the script look for onnxruntime.dll in this temp folder, so is it possible to do so?

Best regards

[fdwr]

@V3nt1n How invested are you in the C++ wrapper? The C++ interface appears to presume static linking and implicitly loads the library for you and retrieves the API via a global singleton upfront Ort::Global<void>::api_, but there's always still the lower-level approach of using the C interface in "onnxruntime_c_api.h", via LoadLibrary, then GetProcAddress for OrtGetApiBase. Alternately, looking at "onnxruntime_cxx_api.h" again, it appears one can now set ORT_API_MANUAL_INIT along with ORT_DLL_IMPORT ("Define ORT_DLL_IMPORT if your program is dynamically linked to Ort.") and manually initialize the primary API pointer yourself. <disclaimer>I haven't tried this myself.</disclaimer>

For example ... SetDllDirectoryW

Prefer LoadLibrary with an explicit path over messing with the DLL search directory list, which can screw up other components when they look for libraries.

[V3nt1n]

Hello fdwr. Thnx for your answer
Finally I was able to run the FMU, I just need to define the path to the files and .dll within the FMU accordingly to the FMU synaxis, and store the files in the proper folders.

• /binaries/win64/onnxruntime.dll
• /resources/model.onnx
• std::string model_path = std::string(fmuResourcesLocation) + "/model.onnx"

Anyway I'll explore your suggestion, because I'm still interested in pack the FMU without the onnx libs inside
Best regards