Building OpenCV for Python with Cuda 12.5 and cuDNN 9.1.1: A dynamic link library (DLL) initialization routine failed.

[dmgardner88]

System Information

Operating System / Platform: Windows 10 Pro
opencv: 4.10.0-pre - 4.x branch up to commit 1db6a8a
opencv_contrib: 4.x branch up to commit 1ed3dd2c
Anaconda3: 24.5.0
Python version: 3.11.7
Cuda: 12.4 or 12.5
Cudnn: 9.1.1
CMake: 3.29.3
Visual Studio 17 2022
Video_Codec_SDK_12.2.72
GPU: NVIDIA GeForce GTX 1070
CPU: Intel Core i5-6600k @3.50GHz

Detailed description

Followed the guidance here on building OpenCV from source. This is with a new, clean install of Windows 10 Pro and the dependencies listed above.

After building from source with Python integration, I'm seeing an error when attempting to import the Python module. The first error is a DLL not found. Following the guidance here and using Process Monitor it looks like the CUDA DLLs were not being found, so I include the Cuda bin directory. After attempting to import again there is an error that a DLL initialization routine failed.

Can't find much online for this error in the context of OpenCV, but this Stackoverflow question mentions that it could be an issue with the CPU not supporting AVX. I've confirmed that this CPU (Intel i5 6600k) support AVX. I also tried building with Cuda 12.4 thinking that it may be due to cudnn being incompatable with Cuda 12.5.

(base) C:\Users\dgard>python
Python 3.11.7 | packaged by Anaconda, Inc. | (main, Dec 15 2023, 18:05:47) [MSC v.1916 64 bit (AMD64)] on win32
Type "help", "copyright", "credits" or "license" for more information.
>>> import cv2
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "C:\Users\dgard\anaconda3\Lib\site-packages\cv2\__init__.py", line 181, in <module>
    bootstrap()
  File "C:\Users\dgard\anaconda3\Lib\site-packages\cv2\__init__.py", line 153, in bootstrap
    native_module = importlib.import_module("cv2")
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\dgard\anaconda3\Lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: DLL load failed while importing cv2: The specified module could not be found.
>>> import os
>>> os.add_dll_directory("C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.5\\bin")
<AddedDllDirectory('C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.5\\bin')>
>>> import cv2
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
ImportError: DLL load failed while importing cv2: A dynamic link library (DLL) initialization routine failed.
>>>

Steps to reproduce

• Install the following, in order:
  • Windows 10 Pro
  • Visual Studio
  • Anaconda3
  • Cuda Toolkit
  • Cudnn
  • CMake
  • Video_Codec_SDK_12.2.72
• Merge the Video Codec SDK Interface and Lib directories into the Cuda v12.5\include and v12.5\lib directories respectively
• Merge the Cudnn \bin\12.4, \include\12.4, and \lib\12.4 directories into Cuda's v12.5\bin, v12.5\lib, v12.5\include directories respectively.
• Run the following CMake command to generate the build configuration:

set CMAKE_BUILD_PARALLEL_LEVEL=4

"C:\Program Files\CMake\bin\cmake.exe" -H"C:\Users\dgard\opencvGPU\opencv" ^
-DOPENCV_EXTRA_MODULES_PATH="C:\Users\dgard\opencvGPU\opencv_contrib\modules" ^
-B"C:\Users\dgard\opencvGPU\build" -G"Visual Studio 17 2022" ^
-DINSTALL_TESTS=ON ^
-DINSTALL_C_EXAMPLES=ON ^
-DBUILD_EXAMPLES=ON ^
-DBUILD_opencv_world=ON ^
-DENABLE_CUDA_FIRST_CLASS_LANGUAGE=ON ^
-DWITH_CUDA=ON ^
-DCUDA_GENERATION=Auto ^
-DBUILD_opencv_python3=ON ^
-DPYTHON3_INCLUDE_DIR=C:/Users/dgard/anaconda3/include ^
-DPYTHON3_LIBRARY=C:/Users/dgard/anaconda3/libs/python311.lib ^
-DPYTHON3_EXECUTABLE=C:/Users/dgard/anaconda3/python.exe ^
-DPYTHON3_NUMPY_INCLUDE_DIRS=C:/Users/dgard/anaconda3/Lib/site-packages/numpy/core/include ^
-DPYTHON3_PACKAGES_PATH=C:/Users/dgard/anaconda3/Lib/site-packages

The CMake configuration for the build was:

-- General configuration for OpenCV 4.10.0-pre =====================================
--   Version control:               4.9.0-546-g1db6a8a1f3
--
--   Extra modules:
--     Location (extra):            C:/Users/dgard/opencvGPU/opencv_contrib/modules
--     Version control (extra):     4.9.0-83-g1ed3dd2c
--
--   Platform:
--     Timestamp:                   2024-06-01T08:06:09Z
--     Host:                        Windows 10.0.19045 AMD64
--     CMake:                       3.29.3
--     CMake generator:             Visual Studio 17 2022
--     CMake build tool:            C:/Program Files/Microsoft Visual Studio/2022/Community/MSBuild/Current/Bin/amd64/MSBuild.exe
--     MSVC:                        1940
--     Configuration:               Debug Release
--
--   CPU/HW features:
--     Baseline:                    SSE SSE2 SSE3
--       requested:                 SSE3
--     Dispatched code generation:  SSE4_1 SSE4_2 FP16 AVX AVX2 AVX512_SKX
--       requested:                 SSE4_1 SSE4_2 AVX FP16 AVX2 AVX512_SKX
--       SSE4_1 (18 files):         + SSSE3 SSE4_1
--       SSE4_2 (2 files):          + SSSE3 SSE4_1 POPCNT SSE4_2
--       FP16 (1 files):            + SSSE3 SSE4_1 POPCNT SSE4_2 FP16 AVX
--       AVX (9 files):             + SSSE3 SSE4_1 POPCNT SSE4_2 AVX
--       AVX2 (38 files):           + SSSE3 SSE4_1 POPCNT SSE4_2 FP16 FMA3 AVX AVX2
--       AVX512_SKX (8 files):      + SSSE3 SSE4_1 POPCNT SSE4_2 FP16 FMA3 AVX AVX2 AVX_512F AVX512_COMMON AVX512_SKX
--
--   C/C++:
--     Built as dynamic libs?:      YES
--     C++ standard:                11
--     C++ Compiler:                C:/Program Files/Microsoft Visual Studio/2022/Community/VC/Tools/MSVC/14.40.33807/bin/Hostx64/x64/cl.exe  (ver 19.40.33811.0)
--     C++ flags (Release):         /DWIN32 /D_WINDOWS /W4 /GR  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi  /fp:precise     /EHa /wd4127 /wd4251 /wd4324 /wd4275 /wd4512 /wd4589 /wd4819 /MP  /O2 /Ob2 /DNDEBUG
--     C++ flags (Debug):           /DWIN32 /D_WINDOWS /W4 /GR  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi  /fp:precise     /EHa /wd4127 /wd4251 /wd4324 /wd4275 /wd4512 /wd4589 /wd4819 /MP  /Zi /Ob0 /Od /RTC1
--     C Compiler:                  C:/Program Files/Microsoft Visual Studio/2022/Community/VC/Tools/MSVC/14.40.33807/bin/Hostx64/x64/cl.exe
--     C flags (Release):           /DWIN32 /D_WINDOWS /W3  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi  /fp:precise     /MP   /O2 /Ob2 /DNDEBUG
--     C flags (Debug):             /DWIN32 /D_WINDOWS /W3  /D _CRT_SECURE_NO_DEPRECATE /D _CRT_NONSTDC_NO_DEPRECATE /D _SCL_SECURE_NO_WARNINGS /Gy /bigobj /Oi  /fp:precise     /MP /Zi /Ob0 /Od /RTC1
--     Linker flags (Release):      /machine:x64  /INCREMENTAL:NO
--     Linker flags (Debug):        /machine:x64  /debug /INCREMENTAL
--     ccache:                      NO
--     Precompiled headers:         NO
--     Extra dependencies:          C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/cudart_static.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppial.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppc.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppitc.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppig.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppist.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppidei.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/cublas.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/cublasLt.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/cufft.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppif.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppim.lib C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v12.5/lib/x64/nppicc.lib
--     3rdparty dependencies:
--
--   OpenCV modules:
--     To be built:                 aruco bgsegm bioinspired calib3d ccalib core cudaarithm cudabgsegm cudacodec cudafeatures2d cudafilters cudaimgproc cudalegacy cudaobjdetect cudaoptflow cudastereo cudawarping cudev datasets dnn dnn_objdetect dnn_superres dpm face features2d flann fuzzy gapi hfs highgui img_hash imgcodecs imgproc intensity_transform line_descriptor mcc ml objdetect optflow phase_unwrapping photo plot python3 quality rapid reg rgbd saliency shape signal stereo stitching structured_light superres surface_matching text tracking ts video videoio videostab wechat_qrcode world xfeatures2d ximgproc xobjdetect xphoto
--     Disabled:                    -
--     Disabled by dependency:      -
--     Unavailable:                 alphamat cannops cvv freetype hdf java julia matlab ovis python2 python2 sfm viz
--     Applications:                tests perf_tests examples apps
--     Documentation:               NO
--     Non-free algorithms:         NO
--
--   Windows RT support:            NO
--
--   GUI:
--     Win32 UI:                    YES
--     VTK support:                 NO
--
--   Media I/O:
--     ZLib:                        build (ver 1.3.1)
--     JPEG:                        build-libjpeg-turbo (ver 3.0.3-70)
--       SIMD Support Request:      YES
--       SIMD Support:              NO
--     WEBP:                        build (ver encoder: 0x020f)
--     PNG:                         build (ver 1.6.43)
--       SIMD Support Request:      YES
--       SIMD Support:              YES (Intel SSE)
--     TIFF:                        build (ver 42 - 4.6.0)
--     JPEG 2000:                   build (ver 2.5.0)
--     OpenEXR:                     build (ver 2.3.0)
--     HDR:                         YES
--     SUNRASTER:                   YES
--     PXM:                         YES
--     PFM:                         YES
--
--   Video I/O:
--     DC1394:                      NO
--     FFMPEG:                      YES (prebuilt binaries)
--       avcodec:                   YES (58.134.100)
--       avformat:                  YES (58.76.100)
--       avutil:                    YES (56.70.100)
--       swscale:                   YES (5.9.100)
--       avresample:                YES (4.0.0)
--     GStreamer:                   NO
--     DirectShow:                  YES
--     Media Foundation:            YES
--       DXVA:                      YES
--
--   Parallel framework:            Concurrency
--
--   Trace:                         YES (with Intel ITT)
--
--   Other third-party libraries:
--     Intel IPP:                   2021.11.0 [2021.11.0]
--            at:                   C:/Users/dgard/opencvGPU/build/3rdparty/ippicv/ippicv_win/icv
--     Intel IPP IW:                sources (2021.11.0)
--               at:                C:/Users/dgard/opencvGPU/build/3rdparty/ippicv/ippicv_win/iw
--     Lapack:                      NO
--     Eigen:                       NO
--     Custom HAL:                  NO
--     Protobuf:                    build (3.19.1)
--     Flatbuffers:                 builtin/3rdparty (23.5.9)
--
--   NVIDIA CUDA:                   YES (ver 12.5.40, CUFFT CUBLAS NVCUVID NVCUVENC)
--     NVIDIA GPU arch:             61
--     NVIDIA PTX archs:
--
--   cuDNN:                         YES (ver 9.1.1)
--
--   OpenCL:                        YES (NVD3D11)
--     Include path:                C:/Users/dgard/opencvGPU/opencv/3rdparty/include/opencl/1.2
--     Link libraries:              Dynamic load
--
--   Python 3:
--     Interpreter:                 C:/Users/dgard/anaconda3/python.exe (ver 3.11.7)
--     Libraries:                   C:/Users/dgard/anaconda3/libs/python311.lib (ver 3.11.7)
--     Limited API:                 NO
--     numpy:                       C:/Users/dgard/anaconda3/Lib/site-packages/numpy/core/include (ver 1.26.4)
--     install path:                C:/Users/dgard/anaconda3/Lib/site-packages/cv2/python-3.11
--
--   Python (for build):            C:/Users/dgard/anaconda3/python.exe
--
--   Java:
--     ant:                         NO
--     Java:                        NO
--     JNI:                         NO
--     Java wrappers:               NO
--     Java tests:                  NO
--
--   Install to:                    C:/Users/dgard/opencvGPU/build/install
-- -----------------------------------------------------------------
--
-- Configuring done (35.9s)
-- Generating done (153.6s)
-- Build files have been written to: C:/Users/dgard/opencvGPU/build

• Run the following CMake command to build

"C:\Program Files\CMake\bin\cmake.exe" --build C:/Users/dgard/opencvGPU/build --target install --config Release

• Validate the install with

"C:\Users\dgard\opencvGPU\build\bin\Release\opencv_test_cudaarithm.exe" --gtest_filter=CUDA_Arithm/GEMM.Accuracy/0

• Attempt to import the cv2 module in an Anaconda Prompt.

Issue submission checklist

• I report the issue, it's not a question
• I checked the problem with documentation, FAQ, open issues, forum.opencv.org, Stack Overflow, etc and have not found any solution
• I updated to the latest OpenCV version and the issue is still there
• There is reproducer code and related data files (videos, images, onnx, etc)

[cudawarped]

I would check that you have the Visual Studio redistributable (I can't see why you wouldn't) and that there are no other installs of OpenCV (through pip or conda) but I suspect its a anaconda issue. Have you tried https://github.com/conda-forge/miniforge/ as suggested in the guide?

If you upload your .pyd and OpenCV dll's I could take a look.

Edit: I just successfuly built and imported the OpenCV python bindings using the latest commits from the 4.x branches against CUDA 12.5, cuDNN 9.1 and miniforge with python 3.10.

[dmgardner88]

Thank you for taking a look. I successfully built and imported the module with the latest 4.x branch commits, Cuda 12.3, cuDNN 8.9.5, the same Anaconda version, but also downgraded MSVC to 19.36 due to CUDA incompatibility with 19.40. Trying to match the build closely to build config in the blog post.

I'm wondering if the original issue is related to the compiler version. I can try building again updating CUDA and cuDNN to the latest versions. If that build works, then can try updating to the latest MSVC version to see if the original issue reproduces, then try removing Anaconda and using miniforge to see if that resolves it. Can provide the files if the issue reproduces.

[cudawarped]

I'm wondering if the original issue is related to the compiler version. I can try building again updating CUDA and cuDNN to the latest versions. If that build works, then can try updating to the latest MSVC version to see if the original issue reproduces, then try removing Anaconda and using miniforge to see if that resolves it. Can provide the files if the issue reproduces.

@dmgardner88 I suspect the issue may be that when you load cv2 it is finding old visual studio redistributable's in your anaconda install.

Try conda update vc14_runtime before import cv2 when building with the newest version of Visual Studio.

[dmgardner88]

Rebuilt with the same Open CV 4.x branch, Cuda 12.5, and cuDNN 9.1.1, and MSVC 19.36 and that build completed and imported successfully.

Then did a build with the same configuration and MSVC 19.40, and that builds but has the originally reported error when importing the module. It reproduces the issue. Here are the OpenCV dlls and the .pyd: opencv.zip

Tried conda update vc14_runtime and got the following error:

(base) C:\Users\dgard>conda update vc14_runtime

PackageNotInstalledError: Package is not installed in prefix.
  prefix: C:\Users\dgard\anaconda3
  package name: vc14_runtime

(base) C:\Users\dgard>conda list
# packages in environment at C:\Users\dgard\anaconda3:
#
# Name                    Version                   Build  Channel
_anaconda_depends         2024.02             py311_mkl_1
...
utf8proc                  2.6.1                h2bbff1b_1
validators                0.18.2             pyhd3eb1b0_0
vc                        14.2                 h21ff451_1
vs2015_runtime            14.27.29016          h5e58377_2
w3lib                     2.1.2           py311haa95532_0

Thanks for the help. Going back to the blog post and matching the build put me on the right track. Will leave the issue open since it is reproducible.

If other folks hit the same issue here's how I worked around it.

You can downgrade MSVC using the visual studio installer from the command line with admin permissions as described here.

"C:\Program Files (x86)\Microsoft Visual Studio\Installer\setup.exe" modify --installPath "C:\Program Files\Microsoft Visual Studio\2022\Community" --add Microsoft.VisualStudio.Component.VC.14.36.17.6.x86.x64 --passive

The component options for install are here.

Instead of copying cuDNN files over to the Cuda directory, I just specified the CMake arguments as follows. I also found I needed to explicit set the CUDA_GENERATION since the auto detection wasn't working with older dependencies.

"C:\Program Files\CMake\bin\cmake.exe" -H"C:\Users\dgard\opencvGPU\opencv" ^
-D OPENCV_EXTRA_MODULES_PATH="C:\Users\dgard\opencvGPU\opencv_contrib\modules" ^
-B"C:\Users\dgard\opencvGPU\build" ^
-G"Visual Studio 17 2022" ^
-D INSTALL_TESTS=ON ^
-D INSTALL_C_EXAMPLES=ON ^
-D BUILD_EXAMPLES=ON ^
-D BUILD_opencv_world=ON ^
-D ENABLE_CUDA_FIRST_CLASS_LANGUAGE=ON ^
-D WITH_CUDA=ON ^
-D CUDA_GENERATION=Pascal ^
-D BUILD_opencv_python3=ON ^
-D PYTHON3_INCLUDE_DIR=C:/Users/dgard/anaconda3/include ^
-D PYTHON3_LIBRARY=C:/Users/dgard/anaconda3/libs/python311.lib ^
-D PYTHON3_EXECUTABLE=C:/Users/dgard/anaconda3/python.exe ^
-D PYTHON3_NUMPY_INCLUDE_DIRS=C:/Users/dgard/anaconda3/Lib/site-packages/numpy/core/include ^
-D PYTHON3_PACKAGES_PATH=C:/Users/dgard/anaconda3/Lib/site-packages ^
-D CUDNN_INCLUDE_DIR="C:\Program Files\NVIDIA\CUDNN\v9.1\include\12.4" ^
-D CUDNN_LIBRARY="C:\Program Files\NVIDIA\CUDNN\v9.1\lib\12.4\x64\cudnn.lib"

I added C:\Program Files\NVIDIA\CUDNN\v9.1\bin to PATH.

Then when importing the python module also added the CUDA and cuDNN directories:

import os
os.add_dll_directory("C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.5\\bin")
os.add_dll_directory("C:\\Program Files\\NVIDIA\\CUDNN\\v9.1\\bin\\12.4")

import cv2

[cudawarped]

Tried conda update vc14_runtime and got the following error:

Maybe conda upgrade vc will work? Either way I suspect the visual c++ redistributable libraries in the root of you anaconda install are being loaded in preference to the ones installed with visual studio. You can check in process monitor, if you get SUCCESS status for any of the resistributable files (e.g. vcruntime140_1.dll) in the root of your python install instead of or in addition to those found in C:\Windows\System32 and they are old (check the file date) then you need to upgrade them.

[evan-kolberg]

#25689 (comment)
Thanks @dmgardner88. Solved it for me.

I needed to specify these locations despite doing this in CMake already:

import os
os.add_dll_directory("C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.5\\bin")
os.add_dll_directory("C:\\Program Files\\NVIDIA\\CUDNN\\v9.1\\bin\\12.4")
import cv2

I took some notes as I was going through the process:

*** MAKE AN ANACONDA ENV ~ this doesn't work with regular python env's; they aren't complete

as of 6/13/24, with cuda 12.5 and cudnn 9.1.1, you need a specific visual studio version or it won't build correctly
https://github.com/opencv/opencv/issues/25689#issuecomment-2151447870

So, install latest visual studio, then use this to downgrade (run cmd as administrator):
"C:\Program Files (x86)\Microsoft Visual Studio\Installer\setup.exe" modify --installPath "C:\Program Files\Microsoft Visual Studio\2022\Community" --add Microsoft.VisualStudio.Component.VC.14.36.17.6.x86.x64 --passive

add opencv and the contrib as git submodules (git submodule add ...)

get cuda toolkit and then get the cuDNN files (using CUDA 12.5 and cuDNN 9.1.1)

run both installers

make sure the locations are in path

open cmake and put in path of source and build folder

check/enter these: ** might have to click configure before finding some cuda options
What to do in cmake: ***NOTE Click on `Advanced` ~ can't see some w/o it on

Source: "D:/programming/fanduel-boggle-cheat/opencv" 
Build folder: "D:\programming\fanduel-boggle-cheat\opencv\build" 
Compiler: "Visual Studio 17 2022" 

-OPENCV_EXTRA_MODULES_PATH="D:/programming/fanduel-boggle-cheat/opencv_contrib/modules"
-PYTHON3_INCLUDE_DIR="C:/Users/evank/anaconda3/envs/opencv_cuda_cudnn/include"
-PYTHON3_LIBRARY="C:/Users/evank/anaconda3/envs/opencv_cuda_cudnn/libs/python311.lib"
-PYTHON3_EXECUTABLE="C:/Users/evank/anaconda3/envs/opencv_cuda_cudnn/python.exe"
-PYTHON3_NUMPY_INCLUDE_DIRS="C:/Users/evank/anaconda3/envs/opencv_cuda_cudnn/Lib/site-packages/numpy/core/include"
-PYTHON3_PACKAGES_PATH="C:/Users/evank/anaconda3/envs/opencv_cuda_cudnn/Lib/site-packages"
-INSTALL_TESTS=ON 
-INSTALL_C_EXAMPLES=ON 
-BUILD_EXAMPLES=ON 
-BUILD_opencv_world=ON 
-WITH_CUDA=ON
-ENABLE_FAST_MATH=ON
-OPENCV_DNN_CUDA=ON
-BUILD_opencv_dnn=ON
-BUILD_opencv_python3=ON
Configure, then (if it detects wrong verison of cuda, replace all 12.1 with 12.5 for ex):
-CUDA_GENERATION=Pascal
-CUDNN_LIBRARY="C:/Program Files/NVIDIA/CUDNN/v9.1/lib/12.4/x64/cudnn.lib"
-ENABLE_CUDA_FIRST_CLASS_LANGUAGE=ON 
-CUDA_FAST_MATH=ON 
-CUDA_ARCH_BIN=7.5
-WITH_NVCUVENC=OFF
-WITH_NVCUVID=OFF
Configure, then:
-CUDNN_INCLUDE_DIR="C:/Program Files/NVIDIA/CUDNN/v9.1/include/12.4"

7.5 is the compute compatability for a 1660 ti

Click on `Configure' - Pray to god there are no errors.

Then click `Generate'

Then click `Open Project'

If Visual Studio prompts you to install some C++ stuff, do it

Then change it from `Debug` to `Release'

Then right-click on the ALL_BUILD file under CMakeTargets and click `Build'

Pray to god there are no errors.

Then right-click on the INSTALL file and click `Build`

Pray to god there are no errors.


========== Build: 601 succeeded, 0 failed, 0 up-to-date, 0 skipped ==========
========== Build completed at 5:33 PM and took 31:58.271 minutes ==========

========== Build: 3 succeeded, 1 failed, 598 up-to-date, 0 skipped ==========
========== Build completed at 8:45 PM and took 01:19.460 minutes ==========

[evan-kolberg]

The err comes back if you install torch with cuda in that same environment.

Traceback (most recent call last):
  File "d:\programming\fanduel-boggle-cheat\opencv_cuda_cudnn.py", line 4, in <module>
    import cv2
  File "C:\Users\evank\anaconda3\envs\opencv_cuda_cudnn\Lib\site-packages\cv2\__init__.py", line 181, in <module>
    bootstrap()
  File "C:\Users\evank\anaconda3\envs\opencv_cuda_cudnn\Lib\site-packages\cv2\__init__.py", line 153, in bootstrap
    native_module = importlib.import_module("cv2")
                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "C:\Users\evank\anaconda3\envs\opencv_cuda_cudnn\Lib\importlib\__init__.py", line 126, in import_module
    return _bootstrap._gcd_import(name[level:], package, level)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ImportError: DLL load failed while importing cv2: The specified module could not be found.

Even when running this code:

import os
os.add_dll_directory("C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.5\\bin")
os.add_dll_directory("C:\\Program Files\\NVIDIA\\CUDNN\\v9.1\\bin\\12.4")
import cv2

count = cv2.cuda.getCudaEnabledDeviceCount()
print(count)

I officially quit. I'm done. This library clearly has too many issues to work properly. Most guides are outdated. I've spent days and days on this. I don't understand why it's so difficult to get this to work. If you want CUDA support in pytorch, you run one install command and it works flawlessly.

[cudawarped]

The err comes back if you install torch with cuda in that same environment.

This is always a possiblity when you mix a package managed install with a custom one. That is the package manager doesn't know anything about the dependencies of your manually built install so it can/will remove add things which can/will brake it.

As you have devoted so much time to this I would advise checking which DLL is missing using process monitor see https://cudawarped.github.io/opencv-experiments/nbs/ImportError_dll_load_failed_while_importing_cv2.html

Note:

1. os.add_dll_directory("C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v12.5\\bin") is unecessary as of cuda: Add missing python CUDA dependency when CUDA is a first class language #25706

2. os.add_dll_directory("C:\\Program Files\\NVIDIA\\CUDNN\\v9.1\\bin\\12.4"), -CUDNN_LIBRARY="C:/Program Files/NVIDIA/CUDNN/v9.1/lib/12.4/x64/cudnn.lib" and -CUDNN_INCLUDE_DIR="C:/Program Files/NVIDIA/CUDNN/v9.1/include/12.4" are unecessary if you download and extract the tarball as I suggested in DNN: CUDA backend requires cuDNN. Please resolve dependency or disable OPENCV_DNN_CUDA=OFF #25426 (comment)

3. So, install latest visual studio, then use this to downgrade (run cmd as administrator):
   "C:\Program Files (x86)\Microsoft Visual Studio\Installer\setup.exe" modify --installPath "C:\Program Files\Microsoft Visual Studio\2022\Community" --add Microsoft.VisualStudio.Component.VC.14.36.17.6.x86.x64 --passive

   Is unecessary if unless you are getting this error

   ImportError: DLL load failed while importing cv2: A dynamic link library (DLL) initialization routine failed

   and even then I would advise upgrading your annaconda vc dependencies as I suggested above.

4. If you install cuDNN into your CUDA toolkit then you could try the following pip wheel once you have removed your existing opencv python installation https://github.com/cudawarped/opencv-python-cuda-wheels/releases/tag/4.10.0.82.

[evan-kolberg]

The pre-built wheel worked!!! Lesson learned: don't try to reinvent the wheel (literally).

I don't even need to add directories, it works like this:

import cv2

count = cv2.cuda.getCudaEnabledDeviceCount()
print(count)

It even works after installing pytorch with CUDA support.

Thank you @cudawarped!