CUDA can't be enabled for current MSVC versions

[Flashtim]

I have been trying to build Viskores on a couple of different Windows 11 machines using MSVC with the Viskores_ENABLE_CUDA option set to true.
While I can generate project files for Visual Studio 2022 with no errors or warnings in Cmake, Viskores fails to be compiled (2693 errors) due to issues with CUDA/Thrust.

These errors mostly repeat for a whole lot of files:

• ""thrust::detail" is ambiguous"
• "a template argument list is not allowed in a declaration of a primary template"
• "class or struct definition is missing"
• "make_projection_symbol: no fundamental base class"
• "'cudafe++' died with status 0xC0000409"

Further, there are issues with CUDA targets, e.g:

"The command ""C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\bin\nvcc.exe" --use-local-env -ccbin "C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64" -x cu -IC:\Viskores\source\viskores\thirdparty\optionparser -IC:\Viskores\source\viskores\thirdparty\diy -IC:\Viskores\source\viskores\thirdparty\lcl\viskoreslcl -IC:\Viskores\source\viskores\thirdparty\loguru -IC:\Viskores\source -IC:\Viskores\build\include -IC:\Viskores\source\viskores\thirdparty\diy\viskoresdiy\include -IC:\Viskores\build\viskores\thirdparty\diy\viskoresdiy\include\viskoresdiy\mpi -I"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v12.4\include" --keep-dir x64\Debug -maxrregcount=0 --machine 64 --compile -cudart static --generate-code=arch=compute_89,code=sm_89 -std=c++17 -Xcudafe=--diag_suppress=1394,--diag_suppress=766 -Xcudafe=--display_error_number --expt-relaxed-constexpr -Xcompiler="/EHsc -Zi -Ob0 -wd4702,-wd4505 /Gy /bigobj" -g -D_WINDOWS -DVISKORESDIY_NO_THREADS -DVISKORESDIY_MPI_AS_LIB -D_SCL_SECURE_NO_WARNINGS -D_CRT_SECURE_NO_WARNINGS -D"CMAKE_INTDIR="Debug"" -Dviskores_cont_EXPORTS -D_WINDLL -D_MBCS -DWIN32 -D_WINDOWS -DVISKORESDIY_NO_THREADS -DVISKORESDIY_MPI_AS_LIB -D_SCL_SECURE_NO_WARNINGS -D_CRT_SECURE_NO_WARNINGS -D"CMAKE_INTDIR="Debug"" -Dviskores_cont_EXPORTS -Xcompiler "/EHsc /W3 /nologo /Od /FS /Zi /RTC1 /MDd /GR" -Xcompiler "/Fdviskores_cont.dir\Debug\vc143.pdb" -o viskores_cont.dir\Debug\ArrayCopyUnknown.obj "C:\Viskores\source\viskores\cont\internal\ArrayCopyUnknown.cxx"" exited with code 9."

This is true for all CUDA Toolkit versions above 12.4 up to the latest 13.0.2.
My guess is, that this is related to this issue:
NVIDIA/cccl#1493

And when using CUDA Toolkit 13.x, additionally, this issue appears:
https://forums.developer.nvidia.com/t/cuda-13-0-missing-thrust-folder/342958

Until now, I switched using GCC within WSL, but I think it would be good to have this resolved.
By the way, the same issue appear when activating Viskores within VTK.
Let me know, if I can provide any more information that might help.
Thanks

---

CMake 4.1.2 (Only the Viskores_ENABLE_CUDA Option was changed from default)
CMake Build for Visual Studio 17 2022
Viskores 1.1.0-RC1.0 (Bit I've tested this for RC2 and older version as well)
Windows 11
MSVC | Microsoft (R) C/C++ Optimizing Compiler Version 19.44.35220 for x64
CUDA Toolkit 12.4 | Cuda compilation tools, release 12.4, V12.4.131

[B3ret]

I hit a similar thing today and can confirm that for Viskores 1.0.0 (the one that ships with vtk 9.5.2), CUDA 12.3 seems to be the latest version that compiles on recent MSVC.

This compiles:
-- The CUDA compiler identification is NVIDIA 12.3.107 with host compiler MSVC 19.38.33145.0

Anything later runs into the errors mentioned above.
I tried all of these CUDA versions: 12.9.1, 12.6.3, 12.4.1

[kmorel]

Does anyone know if this is a windows-specific issue, or if it happens on any platform with CUDA 12.5+?

[Flashtim]

I think, it is a MSVC issue, but I have only used GCC within WSL with various CUDA versions and had no issues.

[kmorel]

@vicentebolea What do you think of this issue? This is tricky to work with because we don't have a CI build that compiles CUDA on Windows. How hard would it be to add a Windows/CUDA build even if we didn't run the testing?

[vicentebolea]

@vicentebolea What do you think of this issue? This is tricky to work with because we don't have a CI build that compiles CUDA on Windows. How hard would it be to add a Windows/CUDA build even if we didn't run the testing?

We did not have CI coverage for windows+cuda even in the last few years of VTK-m times, I suspect that there might be some unknowns due to our cmake configuration been bias toward unix systems. Adding the CI job will be a trivial thing, however, the difficulty how making that build to complete is unknown.

Here is a the commit that disabled the CUDA jobs (circa 2020) b033b28efd9be58b643c93ef17b0dfb5eac167e7

[kmorel]

Can you start by creating the CI job and opening it as a draft PR. From there I could poke at it a bit.

[vicentebolea]

Can you start by creating the CI job and opening it as a draft PR. From there I could poke at it a bit.

I do not have cycles to get to it at the moment, Could you add it as an issue to keep track of the task? We would probably want to use the Frank CI for it (they have windows machines).

[kmorel]

Could you add it as an issue to keep track of the task?

Done: #230

[Flashtim]

When coming across an issue with the current Paraview/ANARI/VisRWTX combination on windows,
I saw that VisRTX might have had a similar issue which was fixed here:

NVIDIA/VisRTX#174

(Note, however, it was only tested until CUDA 13.0.2 and in fact does not work for 13.1. anymore)