Error building from source | Windows 10; torch 1.7.0; CUDA 11.4

[SergeiSamuilov]

After I run python setup.py develop, I get the following error:
log_ninja_enabled.txt
log_ninja_disabled.txt

C:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\ATen/core/ivalue_inl.h(389): warning C4101: 'e': unreferenced local variable
C:\pytorch3d-main\pytorch3d\csrc\sample_pdf\sample_pdf_cpu.cpp(35): error C2146: syntax error: missing ';' before identifier 'output_p'
C:\pytorch3d-main\pytorch3d\csrc\sample_pdf\sample_pdf_cpu.cpp(35): error C2065: 'output_p': undeclared identifier
C:\pytorch3d-main\pytorch3d\csrc\sample_pdf\sample_pdf_cpu.cpp(67): error C2065: 'output_p': undeclared identifier
C:\pytorch3d-main\pytorch3d\csrc\sample_pdf\sample_pdf_cpu.cpp(96): error C2065: 'output_p': undeclared identifier
C:\pytorch3d-main\pytorch3d\csrc\sample_pdf\sample_pdf_cpu.cpp(97): error C2065: 'output_p': undeclared identifier
error: command 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30037\bin\HostX86\x64\cl.exe' failed with exit status 2

Could you please give some clarification on what I'm doing wrong

[SergeiSamuilov]

I also tried building pytorch3d version 0.4.0 as it doesn't have sample_pdf and got new error:

C:/Program Files/NVIDIA GPU Computing Toolkit/CUDA/v11.4/include\thrust/system/cuda/config.h(78): fatal error C1189: #error: The version of CUB in your include path is not compatible with this release of Thrust. CUB is now included in the CUDA Toolkit, so you no longer need to use your own checkout of CUB. Define THRUST_IGNORE_CUB_VERSION_CHECK to ignore this.
error: command 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe' failed with exit status 2

[SergeiSamuilov]

After I deleted environment variable CUB_HOME and built 0.4.0, the error is:

C:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\pybind11\detail/common.h(108): warning C4005: 'HAVE_SNPRINTF': macro redefinition
C:\Users\PC\AppData\Local\Programs\Python\Python38\include\pyerrors.h(315): note: see previous definition of 'HAVE_SNPRINTF'
C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe -c C:\pytorch3d-0.4.0\pytorch3d\csrc\blending\sigmoid_alpha_blend.cu -o build\temp.win-amd64-3.8\Release\pytorch3d-0.4.0\pytorch3d\csrc\blending\sigmoid_alpha_blend.obj -IC:\pytorch3d-0.4.0\pytorch3d\csrc -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\torch\csrc\api\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\TH -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\THC "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include" -IC:\Users\PC\AppData\Local\Programs\Python\Python38\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30037\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" -Xcudafe --diag_suppress=dll_interface_conflict_dllexport_assumed -Xcudafe --diag_suppress=dll_interface_conflict_none_assumed -Xcudafe --diag_suppress=field_without_dll_interface -Xcudafe --diag_suppress=base_class_has_different_dll_interface -Xcompiler /EHsc -Xcompiler /wd4190 -Xcompiler /wd4018 -Xcompiler /wd4275 -Xcompiler /wd4267 -Xcompiler /wd4244 -Xcompiler /wd4251 -Xcompiler /wd4819 -Xcompiler /MD -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -std=c++14 -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 --use-local-env
sigmoid_alpha_blend.cu
C:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\pybind11\detail/common.h(108): warning C4005: 'HAVE_SNPRINTF': macro redefinition
C:\Users\PC\AppData\Local\Programs\Python\Python38\include\pyerrors.h(315): note: see previous definition of 'HAVE_SNPRINTF'
C:\Users\PC\AppData\Local\Programs\Python\Python38\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\PC\AppData\Local\Programs\Python\Python38\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:/pytorch3d-0.4.0/pytorch3d/csrc/blending/sigmoid_alpha_blend.cu".
error: command 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe' failed with exit status 1

[bottler]

Several things going on:

• We haven't tried building at all with CUDA 11.4 yet, as the latest PyTorch (1.9.0) I think is designed to work with CUDA 10.2 and 11.1. PyTorch 1.7.0 which you are using had CUDA 11.0 as its newest version. But your errors don't look related to this.

• The error in your first message looks like a problem with the use of __restrict__ in C++. __restrict__ is non standard and isn't known by MSVC. It may help just to delete the single occurrence of __restrict__ in sample_pdf_cpu.cpp . I actually think it might be a problem on some linux builds too. I plan to remove it from the code.

• setup.py always sets THRUST_IGNORE_CUB_VERSION_CHECK if building with cuda support, so it is odd that you would get an error suggesting setting it.

• I don't understand the pybind problem with sigmoid_alpha_blend. Is that with ninja?

[SergeiSamuilov]

1. Regarding the first message: I tried to build without restrict keyword
   And still got the same error.
   C:\pytorch3d-main\pytorch3d\csrc\sample_pdf\sample_pdf_cpu.cpp(35): error C2146: syntax error: missing ';' before identifier 'output_p'
   sample_pdf_cpu.txt
2. It's with ninja disabled for more verbose output.

[bottler]

1. There's something really weird going on. Line 35 unmodified is understandably a syntax error in MSVC. Removing the __restrict__ should leave line 35 looking like

float* output_p = outputs.data_ptr<float>() + start_batch * n_samples;

In your new version, you are missing the first *, which should be some sort of type error but the syntax would be correct. The syntax error suggests that the compiler is still seeing the old code, not your change.

2. I reckon your cuda setup cannot parse torch.extension.h - it is complaining about our one .cu file which includes it. We used to be careful to exclude it - see 85c396f822 - because there were other problems with it on Linux with old compilers. You may be suffering because of your new cuda (e.g. there may be conditional compilation inside torch 1.7's included pybind version which is responding wrongly to the macros set by the new cuda compiler). It should be possible for you to change that file to exclude it - e.g. by switching torch::PackedTensorAccessor64 to at::PackedTensorAccessor64. We might make this change too.

[SergeiSamuilov]

1. Error after I changed the line:

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe -c C:\pytorch3d-main\pytorch3d\csrc\ball_query\ball_query.cu -o build\temp.win-amd64-3.8\Release\pytorch3d-main\pytorch3d\csrc\ball_query\ball_query.obj -IC:\pytorch3d-main\pytorch3d\csrc -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\torch\csrc\api\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\TH -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\THC "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include" -IC:\Users\PC\AppData\Local\Programs\Python\Python38\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30037\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" -Xcudafe --diag_suppress=dll_interface_conflict_dllexport_assumed -Xcudafe --diag_suppress=dll_interface_conflict_none_assumed -Xcudafe --diag_suppress=field_without_dll_interface -Xcudafe --diag_suppress=base_class_has_different_dll_interface -Xcompiler /EHsc -Xcompiler /wd4190 -Xcompiler /wd4018 -Xcompiler /wd4275 -Xcompiler /wd4267 -Xcompiler /wd4244 -Xcompiler /wd4251 -Xcompiler /wd4819 -Xcompiler /MD -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -std=c++14 -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 --use-local-env
ball_query.cu
C:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\pybind11\detail/common.h(108): warning C4005: 'HAVE_SNPRINTF': macro redefinition
C:\Users\PC\AppData\Local\Programs\Python\Python38\include\pyerrors.h(315): note: see previous definition of 'HAVE_SNPRINTF'
C:\Users\PC\AppData\Local\Programs\Python\Python38\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\PC\AppData\Local\Programs\Python\Python38\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:/pytorch3d-main/pytorch3d/csrc/ball_query/ball_query.cu".
error: command 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe' failed with exit status 1

2. Unfortunately, I can't downgrade CUDA, as there is another extension that requires version >11.1. Switching torch::PackedTensorAccessor64 to at::PackedTensorAccessor64 in sigmoid_alpha_blend.cu yields the same error.

sigmoid_alpha_blend_cpu.txt

C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe -c C:\pytorch3d-0.4.0\pytorch3d\csrc\blending\sigmoid_alpha_blend.cu -o build\temp.win-amd64-3.8\Release\pytorch3d-0.4.0\pytorch3d\csrc\blending\sigmoid_alpha_blend.obj -IC:\pytorch3d-0.4.0\pytorch3d\csrc -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\torch\csrc\api\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\TH -IC:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\THC "-IC:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\include" -IC:\Users\PC\AppData\Local\Programs\Python\Python38\include -IC:\Users\PC\AppData\Local\Programs\Python\Python38\include "-IC:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.29.30037\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" -Xcudafe --diag_suppress=dll_interface_conflict_dllexport_assumed -Xcudafe --diag_suppress=dll_interface_conflict_none_assumed -Xcudafe --diag_suppress=field_without_dll_interface -Xcudafe --diag_suppress=base_class_has_different_dll_interface -Xcompiler /EHsc -Xcompiler /wd4190 -Xcompiler /wd4018 -Xcompiler /wd4275 -Xcompiler /wd4267 -Xcompiler /wd4244 -Xcompiler /wd4251 -Xcompiler /wd4819 -Xcompiler /MD -D__CUDA_NO_HALF_OPERATORS__ -D__CUDA_NO_HALF_CONVERSIONS__ -D__CUDA_NO_BFLOAT16_CONVERSIONS__ -D__CUDA_NO_HALF2_OPERATORS__ --expt-relaxed-constexpr -std=c++14 -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 --use-local-env
sigmoid_alpha_blend.cu
C:\Users\PC\AppData\Local\Programs\Python\Python38\lib\site-packages\torch\include\pybind11\detail/common.h(108): warning C4005: 'HAVE_SNPRINTF': macro redefinition
C:\Users\PC\AppData\Local\Programs\Python\Python38\include\pyerrors.h(315): note: see previous definition of 'HAVE_SNPRINTF'
C:\Users\PC\AppData\Local\Programs\Python\Python38\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\PC\AppData\Local\Programs\Python\Python38\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:/pytorch3d-0.4.0/pytorch3d/csrc/blending/sigmoid_alpha_blend.cu".
error: command 'C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.4\bin\nvcc.exe' failed with exit status 1

[bottler]

The change to sigmoid_alpha_blend.cu should look like this
leaving you with
sigmoid_alpha_blend.cu

[SergeiSamuilov]

Yeah, it worked. Built 0.4.0. Thanks a lot!!!