Claude/fix cuda build linking 01 c ji da qc gv av39a8 c5a jj ep #231
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Updated Python version to 3.11 in workflow.
Updated the GitHub Actions workflow for building nvdiffrast on Windows. Changes include setting up Python 3.11, installing dependencies, and building the nvdiffrast wheel.
- Modified setup.py to support building binary wheels with pre-compiled CUDA extensions - Updated ops.py to load pre-compiled extensions when available (falls back to JIT) - Added GitHub Actions workflow for automated binary wheel building - Includes support for Python 3.11, PyTorch 2.2.2, CUDA 11.8 - Binary wheels eliminate need for MSVC Build Tools at runtime - Backwards compatible with original source wheel behavior
- Replace emoji characters with ASCII text in setup.py - Replace emoji characters in ops.py logging - Replace emoji characters in workflow Python commands - Fixes UnicodeEncodeError in Windows CI environment (cp1252 encoding) - All emojis replaced with [BUILD], [OK], [WARNING] tags
- Delete build-nvdiffrast-windows.yml which builds source wheels - Keep only build_binary_wheel.yml which builds true binary wheels - Prevents confusion about which artifact to download
- Change runner from windows-2019 to windows-2022 - Windows 2019 runners retired as of 2025-06-30 - Windows 2022 also has MSVC pre-installed
- Update cuda-toolkit action from v0.2.14 to v0.2.18 - Disable GitHub and local caching (use-github-cache: false) - Fixes 'Got no files in tool cahce' error - Avoids cache service issues
- MSVC is installed but not in PATH during verification - Build tools will auto-detect and configure MSVC environment - Changed verification to optional check with error handling - Workflow will continue even if cl.exe not found in PATH
- Add explicit 'exit 0' at end of verification script - Check LASTEXITCODE instead of try-catch for cl.exe - Prevents PowerShell from propagating non-zero exit code - Verification step will always succeed now
- Add 'continue-on-error: true' to verification step - Suppress where.exe output with Out-Null - Set ErrorActionPreference to Continue - Workflow will proceed even if verification fails - This is just an informational step, not critical
- CUDA 11.8 doesn't support latest VS 2022 versions (MSVC 14.44) - Add -allow-unsupported-compiler flag to nvcc compiler arguments - This allows compilation with newer MSVC versions - Recommended by CUDA error message for compatibility
- Latest VS 2022 STL requires CUDA 12.4+, but we use CUDA 11.8 - Add _ALLOW_COMPILER_AND_STL_VERSION_MISMATCH define - Bypasses STL1002 version check error - Applied to both cxx and nvcc compiler flags
Compiler flag improvements: - Add /bigobj flag for large object file support - Add _SILENCE_ALL_CXX17_DEPRECATION_WARNINGS - Add --expt-relaxed-constexpr for CUDA constexpr extensions - Add --expt-extended-lambda for modern C++ lambda support - Better organized flags with comments Workflow improvements: - Add MAX_JOBS=2 to prevent OOM during parallel compilation - Add DISTUTILS_USE_SDK=1 for proper SDK detection - Add build log capture with Tee-Object - Add detailed error reporting (last 50 lines on failure) - Add build environment variable display These changes address: 1. CUDA 11.8 + VS 2022 compatibility 2. STL version mismatch 3. Memory constraints during compilation 4. Better debugging for build failures
- Buffer.cpp and other .cpp files use NVDR_CHECK_CUDA_ERROR macro - This macro is only defined when NVDR_TORCH is set - Added /DNVDR_TORCH to cxx_flags (was only in nvcc_flags) - Fixes 'NVDR_CHECK_CUDA_ERROR': identifier not found error
- Add ErrorActionPreference = 'Continue' to prevent stopping on warnings - Change from Tee-Object to direct redirection (> build.log 2>&1) - Capture exit code immediately after build command - Display full build output after completion - Fixes issue where stderr warnings caused premature failure
- Use vswhere to find Visual Studio installation path - Import and activate VS DevShell module - Run Enter-VsDevShell to add cl.exe and MSVC tools to PATH - Verify cl.exe is accessible before building - Fixes 'Cannot find compiler cl.exe in PATH' error from nvcc
- Add cusparse (sparse matrix operations) - Add cublas (linear algebra) - Add thrust (parallel algorithms) - Fixes 'cusparse.h': No such file or directory error - PyTorch requires these libraries for compilation
- Remove sub-packages limitation - Install complete CUDA 11.8 toolkit with all libraries - Ensures cusparse.h, cublas.h, thrust headers are included - Fixes persistent 'cusparse.h not found' error
Add -rdc=true flag to enable relocatable device code compilation. This fixes the issue where RasterImpl.cpp/RasterImpl.cu and texture.cpp/texture.cu were both compiling to the same .obj file names, causing the linker to ignore one and resulting in unresolved symbols. Also add cudadevrt library which is required when using -rdc=true.
PyTorch's CUDAExtension skips compiling .cpp files when a .cu file with the same base name exists. This caused RasterImpl.cpp and texture.cpp to not be compiled, resulting in unresolved symbols. Changes: - Renamed RasterImpl.cpp -> RasterImpl_host.cpp - Renamed texture.cpp -> texture_host.cpp - Updated setup.py to reference renamed files - Removed -rdc=true flag (was causing __cudaRegisterLinkedBinary errors) - Removed cudadevrt library dependency This ensures both CUDA kernel files (.cu) and host implementation files (.cpp) are compiled into separate object files without conflicts.
Initialize p.strideX and p.strideY before using them in the byteOffset calculation. Previously, p.strideX was used at line 327 before being initialized at line 331, causing C4700 compiler warning and potential undefined behavior. This fix moves the stride initialization to before the calculation that uses it.
The nvdiffrast_plugin depends on PyTorch and CUDA DLLs which are not in Windows' DLL search path by default. When torch is imported, it adds its lib directory to the DLL search path, allowing the plugin to find its dependencies. Modified the verification test to import torch before importing the plugin directly. This ensures the DLL search path is properly configured.
Moved the wheel upload step to occur immediately after building, before the test step. This ensures the wheel is preserved even if tests fail, preventing loss of the built artifact. The wheel will now be available for download from GitHub Actions artifacts regardless of test outcome.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
No description provided.