AmmarkoV · reymondmeking-dot · Sep 9, 2026 · Sep 9, 2026
diff --git a/.github/workflows/windows.yml b/.github/workflows/windows.yml
@@ -0,0 +1,32 @@
+name: Windows native build
+
+on:
+  push:
+  pull_request:
+  workflow_dispatch:
+
+permissions:
+  contents: read
+
+jobs:
+  msvc:
+    runs-on: windows-2022
+    steps:
+      - uses: actions/checkout@v4
+      - name: Fetch official OpenCV and GLEW development packages
+        shell: pwsh
+        run: |
+          New-Item -ItemType Directory -Path build/deps -Force | Out-Null
+          Invoke-WebRequest https://github.com/opencv/opencv/releases/download/4.10.0/opencv-4.10.0-windows.exe -OutFile build/deps/opencv.exe
+          7z x build/deps/opencv.exe -obuild/deps -y -bso0 -bsp0
+          if ($LASTEXITCODE -ne 0) { throw 'OpenCV extraction failed' }
+          Invoke-WebRequest https://github.com/nigels-com/glew/releases/download/glew-2.2.0/glew-2.2.0-win32.zip -OutFile build/deps/glew.zip
+          Expand-Archive build/deps/glew.zip build/deps
+      - name: Build all Windows targets and run model-free regression tests
+        shell: pwsh
+        run: |
+          ./scripts/build_windows.ps1 -OpenCVDir build/deps/opencv/build -GLEWRoot build/deps/glew-2.2.0
+          powershell.exe -NoProfile -ExecutionPolicy Bypass -File tests/test_fetch_model.ps1
+          if ($LASTEXITCODE -ne 0) { throw 'PowerShell download tests failed' }
+      # A hosted runner has no guaranteed OpenGL 3.3 desktop. test_wgl_context
+      # is compiled here; opt in to its execution locally with -TestOpenGL.
diff --git a/CMakeLists.txt b/CMakeLists.txt
diff --git a/README.md b/README.md
@@ -292,24 +292,39 @@ CMake handles dependencies automatically:
 > libonnxruntime_providers_cuda.so` or `Could not find an implementation for
 > Expand(13)`, see **[DEPENDENCIES.md](knowledge/DEPENDENCIES.md)** for the cause and fix.
 
-#### Windows (headless build)
-
-Windows is supported as a **headless build only** (MSVC + CMake; OpenCV via
-vcpkg). CMake automatically fetches the `win-x64` ONNX Runtime and configures the
-CLI (`fast_sam_3dbody_run`) and offline BVH extractor (`offline_sam_3dbody_render`).
-
-The live OpenGL overlay viewer (`fast_sam_3dbody_render`) is **not built on
-Windows** — it depends on GLX/X11, which has no in-tree Windows equivalent. CMake
-prints a notice to this effect at configure time. For visualization on Windows,
-use the offline BVH output or the Python frontends. Linux remains the platform
-for live rendering.
-
-Outputs in `build/`:
+#### Windows (native build and WGL viewer)
+
+Native Windows supports the CLI, offline BVH extractor, Python/C shared DLL,
+and an OpenGL 3.3 viewer using Win32/WGL. Use **Visual Studio 2022 x64**, **CMake
+3.21+**, the official **OpenCV 4.10.0** Windows package (`vc16` x64 libraries),
+and **GLEW 2.2.0** for the viewer. From the repository root in PowerShell:
+
+```powershell
+.\scripts\build_windows.ps1 `
+  -OpenCVDir D:\deps\opencv\build `
+  -GLEWRoot D:\deps\glew-2.2.0 -Gpu
+.\tools\fetch_model.ps1 -Profile cuda -Yes
+```
 
-| File | Description |
-|------|-------------|
-| `fast_sam_3dbody_run` | Standalone CLI executable |
-| `libfast_sam_3dbody.so` | Shared library for C++ linking or ctypes |
+`-Gpu` uses GPU ONNX Runtime 1.20.1 without requiring nvcc; without a CUDA
+compiler, native ggml/LBS remains on the CPU. Running the CUDA provider still
+requires CUDA 12.x/cuDNN 9.x runtime DLLs on `PATH`. Use `-Headless` to omit the
+viewer, `-TestOpenGL` to opt into its desktop graphics test, or
+`-OnnxRuntimeDir` to supply an existing ORT package.
+
+| Default Release output | Description |
+|---|---|
+| `build/windows/Release/fast_sam_3dbody_run.exe` | Inference CLI |
+| `build/windows/Release/fast_sam_3dbody_render.exe` | WGL viewer, unless built with `-Headless` |
+| `build/windows/Release/offline_sam_3dbody_render.exe` | Offline BVH extractor |
+| `build/windows/Release/fast_sam_3dbody.dll` | C++/C/Python library with ABI-checked ctypes bindings |
+
+Keep running from the **repository root**: shaders, mesh assets and BVH
+templates retain repository-relative paths. The output folder alone is not a
+standalone distribution. See **[WINDOWS.md](knowledge/WINDOWS.md)** for CPU and
+FP16 model profiles, CUDA-provider commands that do not require TensorRT,
+Python `--lib-dir build/windows`, DLL setup and platform limitations.
+The WGL implementation credits [beemsoft's PR #13](https://github.com/AmmarkoV/SAM3DBody-cpp/pull/13).
 
 ---
 

diff --git a/cmake/StageRuntime.cmake b/cmake/StageRuntime.cmake
@@ -0,0 +1,7 @@
+# A static dependency configuration can legitimately produce an empty list.
+foreach(_dll IN LISTS SAM3D_RUNTIME_DLLS)
+    if(NOT _dll STREQUAL "")
+        get_filename_component(_name "${_dll}" NAME)
+        file(COPY_FILE "${_dll}" "${SAM3D_DESTINATION}/${_name}" ONLY_IF_DIFFERENT)
+    endif()
+endforeach()
diff --git a/knowledge/DEPENDENCIES.md b/knowledge/DEPENDENCIES.md
@@ -10,7 +10,7 @@ at runtime**, you are almost certainly in the right place.
 
 | Component | Version | Notes |
 |-----------|---------|-------|
-| ONNX Runtime | **1.20.1 (GPU build)** | Downloaded automatically by CMake into `build/onnxruntime_dl/` if not found. |
+| ONNX Runtime | **1.20.1 (GPU build)** | Downloaded automatically by CMake into `build/onnxruntime_dl/<full-package-name>/` if not found; for Linux GPU this is `onnxruntime-linux-x64-gpu-1.20.1`. CPU/GPU packages use separate directories. |
 | CUDA | **12.x** | Required by the ORT 1.20.1 CUDA execution provider. |
 | cuDNN | **9.x** | Required by ORT ≥ 1.19. **cuDNN 8 will not work** and is the #1 cause of the CUDA EP failing to load. |
 | NVIDIA driver | Recent enough for CUDA 12 (≥ 525) | Check with `nvidia-smi`. |
@@ -41,7 +41,7 @@ version — virtually always **cuDNN 9** or a CUDA 12 runtime lib.
 nvidia-smi
 
 # 2. THE decisive command — what is the provider .so actually missing?
-ldd build/onnxruntime_dl/lib/libonnxruntime_providers_cuda.so | grep -i "not found"
+ldd build/onnxruntime_dl/onnxruntime-linux-x64-gpu-1.20.1/lib/libonnxruntime_providers_cuda.so | grep -i "not found"
 
 # 3. Is cuDNN 9 / cuBLAS installed and visible to the loader?
 ldconfig -p | grep -i "cudnn\|cublas\|cufft"
@@ -72,7 +72,7 @@ for m in (nvidia.cudnn, nvidia.cublas)))"):$LD_LIBRARY_PATH
 
 ```bash
 # Should now print nothing (no missing libraries):
-ldd build/onnxruntime_dl/lib/libonnxruntime_providers_cuda.so | grep -i "not found"
+ldd build/onnxruntime_dl/onnxruntime-linux-x64-gpu-1.20.1/lib/libonnxruntime_providers_cuda.so | grep -i "not found"
 ```
 
 Re-run `scripts/webcam.sh` — the CUDA EP should load and the pipeline should run
@@ -172,7 +172,7 @@ A system `.deb`/`.tar` TensorRT 10.4 install from developer.nvidia.com/tensorrt
 (then `sudo ldconfig`) works too. Verify the runtime is visible:
 
 ```bash
-ldd build/onnxruntime_dl/lib/libonnxruntime_providers_tensorrt.so | grep -i nvinfer
+ldd build/onnxruntime_dl/onnxruntime-linux-x64-gpu-1.20.1/lib/libonnxruntime_providers_tensorrt.so | grep -i nvinfer
 # every libnvinfer*.so.10 / libnvonnxparser.so.10 line should resolve (no "not found")
 ```