DLL load failed while importing _Ax: The specified module could not be found.

[Ati1984ch]

Expected Behavior

Actual Behavior

Code to reproduce the problem (If applicable)

Specifications

• MATLAB/python version:
• OS:
• CUDA version:

[Ati1984ch]

Hello all,
I hope I have not created a repeated issue here. But I get the built successfully done (Python). Yet when I try the demo3 as a test to see if I can import tigre finally, I get the above error. (Python 3.9)
Appreciate any clues. -I changed the Python version to 3.12 got the same error.

[AnderBiguri]

Heya! This seems to be caused by a recent issue that I still don't know how to generallize.

However, if you follow the instructions in this comment: #628 (comment)

You should be able to compile TIGRE correctly, hopefully! Let me know

[NGao-07]

You can fix this by explicitly adding the CUDA bin directory to the DLL search path in tigre/__init__.py.

Modify your site-packages/tigre/__init__.py as follows:

import os
import platform

# Add CUDA DLL directory for Python 3.8+ on Windows
if platform.system() == "Windows" and hasattr(os, "add_dll_directory"):
    cuda_path = os.environ.get("CUDA_PATH")
    if cuda_path:
        cuda_bin = os.path.join(cuda_path, "bin")
        if os.path.exists(cuda_bin):
            os.add_dll_directory(cuda_bin)
    else:
        # Fallback to default installation path if CUDA_PATH is not set
        default_cuda_bin = r"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8\bin"
        if os.path.exists(default_cuda_bin):
            os.add_dll_directory(default_cuda_bin)

[AnderBiguri]

@NGao-07 This works, but ultimately probably conflicts with #579 . There is an issue here of people installing CUDA normally or via conda. I need some piece of code to make sure it work sin both cases, but unsure what that would be. I think it would be something like checking if there is already a cuda path given by conda first, and otherwise doing this (or the commented code above)

[NGao-07]

@NGao-07 This works, but ultimately probably conflicts with #579 . There is an issue here of people installing CUDA normally or via conda. I need some piece of code to make sure it work sin both cases, but unsure what that would be. I think it would be something like checking if there is already a cuda path given by conda first, and otherwise doing this (or the commented code above)

Thank you.