Link error, Libtorch 1.5 on Windows

[slaweks17]

🐛 Bug

Large number of link errors when trying rebuild an existing program with the new 1.5 version of Libtorch on Windows

To Reproduce

Steps to reproduce the behavior:

1. Create a following program

#include <torch/torch.h>
int main(int argc, char** argv) {
torch::Tensor x_t;
}

2. Try to build it. The compilation is successful, the link fails with messages like
   LNK2019 unresolved external symbol "__declspec(dllimport) public: __cdecl at::Tensor::Tensor(void)" (_imp??0Tensor@at@@qeaa@XZ) referenced in function main Pytorch15 C:\progs\c++\CTorch\Pytorch15\test.obj 1

Environment

Windows 10, Visual Studio 2019, toolset either v141 or v142, the linker options:
/OUT:"C:\progs\c++\CTorch\x64\Debug\Pytorch15.exe" /MANIFEST /NXCOMPAT /PDB:"C:\progs\c++\CTorch\x64\Debug\Pytorch15.pdb" /DYNAMICBASE "torch.lib" "c10.lib" "kernel32.lib" "user32.lib" "gdi32.lib" "winspool.lib" "comdlg32.lib" "advapi32.lib" "shell32.lib" "ole32.lib" "oleaut32.lib" "uuid.lib" "odbc32.lib" "odbccp32.lib" /DEBUG /MACHINE:X64 /INCREMENTAL /PGD:"C:\progs\c++\CTorch\x64\Debug\Pytorch15.pgd" /SUBSYSTEM:CONSOLE /MANIFESTUAC:"level='asInvoker' uiAccess='false'" /ManifestFile:"x64\Debug\Pytorch15.exe.intermediate.manifest" /ERRORREPORT:PROMPT /NOLOGO /LIBPATH:"C:\progs2\libtorch_debug\lib" /TLBID:1

The program builds and works when liked with libtorch 1.4.

[peterjc123]

Well, it looks like you are constructing the VS project by hand. Please link against torch_cpu.lib instead of torch.lib. If you are using CUDA, then you'll also need link against torch_cuda.lib and add the following argument -INCLUDE:?warp_size@cuda@at@@YAHXZ.

[slaweks17]

Great! Thank you very much!!

[jianyin2016]

Well, it looks like you are constructing the VS project by hand. Please link against torch_cpu.lib instead of torch.lib. If you are using CUDA, then you'll need link against torch_cuda.lib and add the following argument -INCLUDE:?warp_size@cuda@at@@YAHXZ.

That is it! but where should I add the argument -INCLUDE:?warp_size@cuda@at@@YAHXZ ?

[peterjc123]

http://nanapro.org/en-us/documentation/img/linker_cmdline.png

[jianyin2016]

That is very nice of you. @peterjc123
thanks!

[xpngzhng]

hi @peterjc123
Sorry to bother you again, I did not understand your solution.

What the the connection between
"__declspec(dllimport) public: __cdecl at::Tensor::Tensor(void)" (_imp??0Tensor@at@@qeaa@XZ)
and
-INCLUDE:?warp_size@cuda@at@@yahxz

for my project, If I only link against torch_cuda.lib, there are 81 unsolved references.
adding -INCLUDE:?warp_size@cuda@at@@yahxz in the Linker -> Command Line -> Additional Options did not work.

[peterjc123]

@xpngzhng For all the linker flags, please refer to the script here. I don't expect some many of you guys creating a VS project by yourself without using CMake.

[xpngzhng]

@peterjc123
got it, thanks

[yoshihingis]

hi @peterjc123
Sorry to bother you again ,
I built mnist.cpp with the libtorch ver1.5 according with your advice , but I did not get good result.

(1) About Linked Libs
torch.lib : not linked.
torch_cpu.lib : linked.
torch_cuda.lib : linked

(2)Commnad line
a. I did not add the additional command , mnist operated correctly at CPU mode.
b. I add the additional command -INCLUDE:?warp_size@cuda@at@@yahxz , GPU was detected and The Message "CUDA available ! Trining on GPU" was indicated, but any information of training progress was never indicated. I attached the capture image

・I made mnist.sln file by cmake by using CmakeLists.txt as blow. I could build mnist with no error.
CMakeLists.txt

But the result was same at GPU MODE, only "CUDA available ! Training on GPU" was indicated ,but
any information of training progress was never indicated.

I want to know your opinion.

Regards,

[peterjc123]

@yoshihingis You'll need to copy the libtorch DLLs to the directory of your target executable.

[yoshihingis]

Dear peterjc123

Thank you for your reply.
But I copy the libtorch Dlls to the same directory of executable file. hence the mnist operated correctly at CPU mode . Unfortunately I faced on the trouble at only GPU mode.

[peterjc123]

@yoshihingis Please try the following commands in CMD.

git clone https://github.com/pytorch/examples.git
curl -kLO https://download.pytorch.org/libtorch/cu102/libtorch-win-shared-with-deps-1.5.0.zip
set LIBTORCH_ROOT=%CD%\libtorch
7z x libtorch-win-shared-with-deps-1.5.0.zip
cd examples\cpp\mnist
mkdir build
cd build
cmake -G"Visual Studio 16 2019" -Ax64 -Thost=x64 -DCMAKE_PREFIX_PATH="%LIBTORCH_ROOT%" ..
cmake --build . --config Release
:: Please note that the directory `data` must be in the current working directory
:: Otherwise, please specify the absolute path in your code
Release\mnist.exe