[dist] Build module with CUDA enabled

Author: Breakthrough

.github/workflows/build_wheels_windows.yml

@@ -1,4 +1,13 @@
-name: Windows x86_64
+# GitHub Actions Workflow: Build opencv-python with CUDA support on Windows
+#
+# This workflow compiles opencv-python from source with CUDA enabled on a
+# GitHub-hosted Windows runner. The resulting Python wheel is then uploaded
+# as a build artifact.
+#
+# This is a complex and long-running process. It is configured to run only
+# on manual trigger (workflow_dispatch).
+
+name: Windows CUDA x64
 
 on:
   workflow_dispatch:
@@ -48,6 +57,54 @@ jobs:
       uses: microsoft/setup-msbuild@v1.1
     - name: Setup NASM
       uses: ilammy/setup-nasm@v1
+    - name: Cache CUDA Toolkit Installer
+      id: cache-cuda-installer
+      uses: actions/cache@v3
+      with:
+        path: ./.cache/cuda_installer.exe
+        key: cuda-installer-12.0.0
+    - name: 🔧 Install NVIDIA CUDA Toolkit
+      run: |
+        $installer_path = "./.cache/cuda_installer.exe"
+        if (-not (Test-Path $installer_path)) {
+          echo "Downloading CUDA Toolkit..."
+          $cuda_installer_url = "https://developer.download.nvidia.com/compute/cuda/12.0.0/network_installers/cuda_12.0.0_windows_network.exe"
+          New-Item -Path (Split-Path $installer_path) -ItemType Directory -Force
+          curl -L -o $installer_path $cuda_installer_url
+        } else {
+          echo "CUDA Toolkit installer found in cache."
+        }
+        echo "Installing CUDA Toolkit silently..."
+        $arguments = "-s -n"
+        Start-Process -FilePath $installer_path -ArgumentList $arguments -Wait -NoNewWindow
+        echo "Adding CUDA to PATH..."
+        $CUDA_PATH = "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.0"
+        echo "CUDA_PATH=$CUDA_PATH" | Out-File -FilePath $env:GITHUB_ENV -Append
+      shell: pwsh
+    - name: Cache CuDNN Installer
+      id: cache-cudnn-installer
+      uses: actions/cache@v3
+      with:
+        path: ./.cache/cudnn.zip
+        key: cudnn-8.9.7.29-cuda-12-windows
+    - name: 🔧 Install NVIDIA CuDNN
+      run: |
+        $cudnn_path = "./.cache/cudnn.zip"
+        if (-not (Test-Path $cudnn_path)) {
+          echo "Downloading CuDNN..."
+          $cudnn_installer_url = "https://developer.download.nvidia.com/compute/cudnn/redist/cudnn/windows-x86_64/cudnn-windows-x86_64-8.9.7.29_cuda12-archive.zip"
+          New-Item -Path (Split-Path $cudnn_path) -ItemType Directory -Force
+          curl -L -o $cudnn_path $cudnn_installer_url
+        } else {
+          echo "CuDNN installer found in cache."
+        }
+        echo "Installing CuDNN..."
+        7z x $cudnn_path
+        $CUDNN_PATH = "/a/opencv-python-cuda/opencv-python-cuda/cudnn-windows-x86_64-8.9.7.29_cuda12-archive"
+        echo "CUDNN_LIBRARY=$CUDNN_PATH/lib/x64" | Out-File -FilePath $env:GITHUB_ENV -Append
+        echo "CUDNN_INCLUDE_DIR=$CUDNN_PATH/include" | Out-File -FilePath $env:GITHUB_ENV -Append
+      shell: pwsh
+
     - name: Build a package
       # CMake 3.25 regression fix. See https://stackoverflow.com/questions/74162633/problem-compiling-from-source-opencv-with-mvsc2019-in-64-bit-version
       run: |

setup.py

@@ -86,8 +86,8 @@ def main():
     if build_headless and not build_contrib:
         package_name = "opencv-python-headless"
 
-    if build_rolling:
-        package_name += "-rolling"
+    #if build_rolling:
+    #    package_name += "-rolling"
 
     package_name = os.environ.get('OPENCV_PYTHON_PACKAGE_NAME', package_name)
 
@@ -162,18 +162,23 @@ def main():
         else ["-G", "Unix Makefiles"]
     )
 
+    cudnn_library = os.environ["CUDNN_LIBRARY"]
+    cudnn_include_dir = os.environ["CUDNN_INCLUDE_DIR"]
+    cuda_arch_bin = "6.0;6.1;7.0;7.5"
+    cuda_arch_ptx = "7.5"
+
     cmake_args = (
         (ci_cmake_generator if is_CI_build else [])
         + [
             # skbuild inserts PYTHON_* vars. That doesn't satisfy opencv build scripts in case of Py3
-            "-DPYTHON3_EXECUTABLE=%s" % sys.executable,
-            "-DPYTHON_DEFAULT_EXECUTABLE=%s" % sys.executable,
-            "-DPYTHON3_INCLUDE_DIR=%s" % python_include_dir,
-            "-DPYTHON3_LIBRARY=%s" % python_lib_path,
+            f"-DPYTHON3_EXECUTABLE={sys.executable}",
+            f"-DPYTHON_DEFAULT_EXECUTABLE={sys.executable}",
+            f"-DPYTHON3_INCLUDE_DIR=\"{python_include_dir}\"",
+            f"-DPYTHON3_LIBRARY=\"{python_lib_path}\"",
             "-DBUILD_opencv_python3=ON",
             "-DBUILD_opencv_python2=OFF",
             # Disable the Java build by default as it is not needed
-            "-DBUILD_opencv_java=%s" % build_java,
+            f"-DBUILD_opencv_java={build_java}",
             # Relative dir to install the built module to in the build tree.
             # The default is generated from sysconfig, we'd rather have a constant for simplicity
             "-DOPENCV_PYTHON3_INSTALL_PATH=python",
@@ -189,6 +194,13 @@ def main():
             "-DBUILD_DOCS=OFF",
             "-DPYTHON3_LIMITED_API=ON",
             "-DBUILD_OPENEXR=ON",
+            "-DWITH_CUDA=ON",
+            "-DWITH_NVCUVID=OFF",
+            f"-DCUDA_ARCH_BIN=\"{cuda_arch_bin}\"",
+            f"-DCUDA_ARCH_PTX=\"{cuda_arch_ptx}\"",
+            "-DOPENCV_ENABLE_NONFREE=ON",
+            f"-DCUDNN_LIBRARY=\"{cudnn_library}\"",
+            f"-DCUDNN_INCLUDE_DIR=\"{cudnn_include_dir}\"",
         ]
         + (
             # CMake flags for windows/arm64 build