Unresolved externals when using thrust

Accelerated Computing CUDA CUDA Programming and Performance
1

Using thrust functions is giving unresolved externals in all of them. I already included

cuda/include, libcudacxx/include, thrust, cub which are all the includes in the samples. I also linked with cudart.lib, cudart_static.lib. cudadevrt.lib. I don’t know what else to include?

In the files I included all the .h files where these functions are, e.g. transform_iterator.h, copy.h, etc.

here are the errors:

unresolved external symbol “class thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::transform_iterator<struct get_color_ref,class thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::detail::normal_iterator,struct thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::use_default,struct thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::use_default> __cdecl thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::cuda_cub::copy<struct thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::cuda_cub::tag,class thrust::THRUST_300104_SM___CUDA_ARCH_LIST___NS::detail::normal_iterator,class

2

Thrust is a header-only library and should not require any linking. Please show a minimal reproducer code and the used compilation command.

I believe your compiler setup is not correct. When correctly compiled with NVCC, the macro __CUDA_ARCH_LIST__ should be expanded to a list of gpu architecture.

For example, this code

#include <thrust/device_vector.h>
#include <iostream>

int main(){
    std::cout << typeid(thrust::device_vector<int>{}).name() << "\n";
}

prints N6thrust16_V_300300_SM_90013device_vectorIiNS0_16device_allocatorIiEEEE when compiled with nvcc -arch=sm_90 , and N6thrust16_V_300300_SM_89013device_vectorIiNS0_16device_allocatorIiEEE when compiled with nvcc -arch=sm_89

(note __CUDA_ARCH_LIST__ was expanded to 900 and 890, respectively)

3

Thanks for the reply. I have been having all kinds of problems with Cuda code which lends me to agree with you that the cuda compiler might be the problem.

Here is how I am using thrust:

	thrust::device_vector<Vertex> d_vertices = m_singleFrameImages[i]->getImageVertices();

	// copy the image pixels to a device vector, use the constructor do not copy
	thrust::device_vector<uchar> d_pixels(m_denoisedImagesPixels[i].get(), m_denoisedImagesPixels[i].get() + (imagesWidth * imagesHeight));

	// Create a fancy iterator that points to the 'color' members of d_vertices
	thrust::transform_iterator<get_color_ref, thrust::device_vector<Vertex>::iterator> x_iter(d_vertices.begin(), get_color_ref());

	// Copy the values from d_pixels into the 'color' members using thrust::copy
	thrust::copy(d_pixels.begin(), d_pixels.end(), x_iter);

I am setting a value in a vector of structs from another vector by using a function (get_color_vector) to do the copying.

I am using Visual studio so I am not compiling on the command line. Here is the command from the log file:

HostCommandLineTemplate = -Xcompiler “/EHsc [Warning] /nologo [Optimization] /FS [DebugInformationFormat] [RuntimeChecks] [Runtime] [TypeInfo]” -Xcompiler “/Fd[ProgramDataBaseFileName]”
SplitCompile = Default
CodeGeneration = compute_86,sm_86;compute_89,sm_89;compute_120,sm_120
AdditionalCompilerOptions =
EnableVirtualArchInFatbin = true
RuntimeApiCommandLineTemplate = -ccbin “C:\Program Files\Microsoft Visual Studio\2022\Community\VC\Tools\MSVC\14.44.35207\bin\HostX64\x64” -x cu [GenerateRelocatableDeviceCode] [ExtensibleWholeProgramCompilation] [Include] [RequiredIncludes] [InterleaveSourceInPTX] [GPUDebugInfo] [GenerateLineInfo] [Keep] [KeepDir] [FastMath] [MaxRegCount] [PtxAsOptionV] [SplitCompile] [FastCompile] [TargetMachinePlatform] [NvccCompilation] [CudaRuntime] [AdditionalOptions] [HostDebugInfo] [Emulation] [Defines] -Xcompiler “/EHsc [Warning] /nologo [Optimization] /FS [DebugInformationFormat] [RuntimeChecks] [Runtime] [TypeInfo]” -Xcompiler “/Fd[ProgramDataBaseFileName]” [CompileOut] “%(FullPath)”

BuildCommandLineTemplate = --use-local-env
GenerateRelocatableDeviceCode = true
UseHostDefines = true
GPUDebugInfo = true
Xcompiler=“/EHsc -Zi -Ob0”
Defines = ;_WINDOWS;CMAKE_INTDIR=“Debug”;Zc:preprocessor
CudaRuntime = Static
NvccCompilation = compile
PropsCacheOutputFile = %(Filename)%(Extension).cache
TargetMachinePlatform = 64
CommandLineTemplate =

               # Driver API (NVCC Compilation Type is .cubin, .gpu, or .ptx)
               set CUDAFE_FLAGS=--sdk_dir "C:\Program Files (x86)\Windows Kits\10\"
               "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\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 [GenerateRelocatableDeviceCode] [ExtensibleWholeProgramCompilation] [Include] [RequiredIncludes] [InterleaveSourceInPTX] [GPUDebugInfo] [GenerateLineInfo] [Keep] [KeepDir] [FastMath] [MaxRegCount] [PtxAsOptionV] [SplitCompile] [FastCompile] [TargetMachinePlatform] [NvccCompilation] [CudaRuntime] [AdditionalOptions] [Defines] [CompileOut] "%(FullPath)"
               
               # Runtime API (NVCC Compilation Type is hybrid object or .c file)
               set CUDAFE_FLAGS=--sdk_dir "C:\Program Files (x86)\Windows Kits\10\"
               "C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\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 [GenerateRelocatableDeviceCode] [ExtensibleWholeProgramCompilation] [Include] [RequiredIncludes] [InterleaveSourceInPTX] [GPUDebugInfo] [GenerateLineInfo] [Keep] [KeepDir] [FastMath] [MaxRegCount] [PtxAsOptionV] [SplitCompile] [FastCompile] [TargetMachinePlatform] [NvccCompilation] [CudaRuntime] [AdditionalOptions] [HostDebugInfo] [Emulation] [Defines] -Xcompiler "/EHsc [Warning] /nologo [Optimization] /FS [DebugInformationFormat] [RuntimeChecks] [Runtime] [TypeInfo]" -Xcompiler "/Fd[ProgramDataBaseFileName]" [CompileOut] "%(FullPath)"

Here is the command line from the linker page:

“C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin\nvcc.exe” -dlink --dlink-time-opt -o MRIViewer\x64\Debug\MRIViewer.device-link.obj -Xcompiler “/EHsc /nologo /Zi " -Xcompiler “/Fd[ProgramDataBaseFileName]” -L"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\bin\x64/crt” -L"C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v13.1\lib\x64" -G

If I compile each file in the project separately I do not get any errors, but when I compile the project it generates the unresolved external error.

does that help show if the problems are coming from the cuda compiler?

Thanks

4

You could try compiling without device debugging (remove flag -G)

You could try compiling the small cuda code I posted above in a fresh project.

Also, the CUDA installation guide for windows recommends verifying the CUDA / VS setup using the provided cuda samples. They come with VS solution files, and the document also explains how to add cuda configuration to an existing project.

I don’t have more suggestions. I do not develop on windows.

5

Thank you for your help. I appreciate it.