Windows 2019 vcc build

[spyroot]

Hi Folks,

I manage to narrow down most of the issues with VC. ( note I keep testing CUDA 11.3 / 11.4 Torch 1.9/1.10)
but it same issue.

I added /wd4624 torch extension so it won't go crazy with false positives.
so it invoked with /wd4267

Tested both with and without 14/17 ( default one wrong it not valid vc args.

    extra_compile_args = {"cxx": ["/std:c++17"]}

also, this kEpsilon is defined in different places. So it is not clear it float or double
In util geometric it just cost auto

in another spot, it is defined below. So it truncates type or can't do type deduction in templates.

I refactored a bit header file in ball etc and used the same epsilon definition,
it looks like fixed all truncation errors.

C:\Users\neros.conda\envs\tf-gpu\lib\site-packages\torch\include\pybind11\cast.h(1429): error: too few arguments for template template parameter "Tuple"
detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
(1507): here

#ifdef _MSC_VER
#define kEpsilon 1e-8f
#else
const auto kEpsilon = 1e-8;
#endif

But I can't narrow down errors related to templates.

FAILED: C:/dev/pytorch3d/build/temp.win-amd64-3.9/Release/dev/pytorch3d/pytorch3d/csrc/ball_query/ball_query.obj
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc --generate-dependencies-with-compile --dependency-output C:\dev\pytorch3d\build\temp.win-amd64-3.9\Release\dev\pytorch3d\pytorch3d\csrc\ball_query\ball_query.obj.d --use-local-env -Xcompiler /MD -Xcompiler /wd4819 -Xcompiler /wd4251 -Xcompiler /wd4244 -Xcompiler /wd4267 -Xcompiler /wd4275 -Xcompiler /wd4018 -Xcompiler /wd4190 -Xcompiler /EHsc -Xcompiler /wd4624 -Xcudafe --diag_suppress=base_class_has_different_dll_interface -Xcudafe --diag_suppress=field_without_dll_interface -Xcudafe --diag_suppress=dll_interface_conflict_none_assumed -Xcudafe --diag_suppress=dll_interface_conflict_dllexport_assumed -DWITH_CUDA -DTHRUST_IGNORE_CUB_VERSION_CHECK -IC:\dev\pytorch3d\pytorch3d\csrc -IC:\dev\cub -IC:\Users\neros\.conda\envs\tf-gpu\lib\site-packages\torch\include -IC:\Users\neros\.conda\envs\tf-gpu\lib\site-packages\torch\include\torch\csrc\api\include -IC:\Users\neros\.conda\envs\tf-gpu\lib\site-packages\torch\include\TH -IC:\Users\neros\.conda\envs\tf-gpu\lib\site-packages\torch\include\THC "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include" -IC:\Users\neros\.conda\envs\tf-gpu\include -IC:\Users\neros\.conda\envs\tf-gpu\include -ID:\dev\VC\Tools\MSVC\14.29.30133\ATLMFC\include -ID:\dev\VC\Tools\MSVC\14.29.30133\include "-IC:\Program Files (x86)\Windows Kits\NETFXSDK\4.8\include\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.19041.0\ucrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.19041.0\shared" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.19041.0\um" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.19041.0\winrt" "-IC:\Program Files (x86)\Windows Kits\10\include\10.0.19041.0\cppwinrt" -c C:\dev\pytorch3d\pytorch3d\csrc\ball_query\ball_query.cu -o C:\dev\pytorch3d\build\temp.win-amd64-3.9\Release\dev\pytorch3d\pytorch3d\csrc\ball_query\ball_query.obj -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -DCUDA_HAS_FP16=1 -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=_C -D_GLIBCXX_USE_CXX11_ABI=0 -gencode=arch=compute_86,code=compute_86 -gencode=arch=compute_86,code=sm_86
C:\Users\neros\.conda\envs\tf-gpu\lib\site-packages\torch\include\pybind11\cast.h(1429): error: too few arguments for template template parameter "Tuple"
          detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
(1507): here

C:\Users\neros\.conda\envs\tf-gpu\lib\site-packages\torch\include\pybind11\cast.h(1503): error: too few arguments for template template parameter "Tuple"
          detected during instantiation of class "pybind11::detail::tuple_caster<Tuple, Ts...> [with Tuple=std::pair, Ts=<T1, T2>]"
(1507): here

2 errors detected in the compilation of "c:/dev/pytorch3d/pytorch3d/csrc/ball_query/ball_query.cu".
ball_query.cu

I'm not even sure how I can dump errors because it is not clear what error actually even is.

[bottler]

If you are using the latest main, there is no problem with the code. Multiple people have it working. Don't try to fix the compiler errors by changing the code. Are you using visual studio 16.9.6?

[spyroot]

Thank you for your feedback !, No it is Version 16.11.5 Torch 1.10 / 1.19 Cuda 11.3 / 11.4. it strange really I re-create 20 times different conda env.

[bottler]

I think you may need to downgrade the compiler. See #876 (comment) and the links from there. I think it's a known problem with the very latest compiler.

[spyroot]

So it took some effort it looks like the issue somewhere in Torch 1.19/1.10 and CUDA 11.4. I manage to compile and testing now.