How can I set -arch=sm_NN of nvcc in my CMake configuration?

I have a .cu file which uses atomicCAS with inputs (unsigned short *, unsigned short, unsigned short). My environment is with --gpu-code=sm_86,sm_86,sm_61. It occurs compile error under the default cuda architecture.

I tested using nvcc -arch=sm_86 my.cu, this funciton could be compiled successfully. So, how could I set -arch=sm_86 of nvcc in my CMakeLists.txt?

In CMake 3.18 and above, you do this by setting the architecture numbers in the CUDA_ARCHITECTURES target property (which is default initialized according to the CMAKE_CUDA_ARCHITECTURES variable) to a semicolon separated list (CMake uses semicolons as its list entry separator character).

From the docs' Examples section:

> cmake
> set_target_properties(tgt PROPERTIES CUDA_ARCHITECTURES "35;50;72")
>
> Generates code for real and virtual architectures 30, 50 and 72.
>
> cmake
> set_property(TARGET tgt PROPERTY CUDA_ARCHITECTURES 70-real 72-virtual)
>
> Generates code for real architecture 70 and virtual architecture 72.
>
> cmake
> set_property(TARGET tgt PROPERTY CUDA_ARCHITECTURES OFF)
>
> CMake will not pass any architecture flags to the compiler.

The source code that sets the build command with the sm_ argument for the Makefile generator can be found in Source/cmMakefileTargetGenerator.cxx:

void cmMakefileTargetGenerator::WriteDeviceLinkRule(
  std::vector<std::string>& commands, const std::string& output)
{
  std::string architecturesStr =
    this->GeneratorTarget->GetSafeProperty("CUDA_ARCHITECTURES");
  ...
  std::vector<std::string> architectures = cmExpandedList(architecturesStr);
  ...
  for (const std::string& architectureKind : architectures) {
    ...
    const std::string architecture =
      architectureKind.substr(0, architectureKind.find('-'));
    ...
    std::string command = cmStrCat(
      this->Makefile->GetRequiredDefinition("CMAKE_CUDA_DEVICE_LINKER"),
      " -arch=sm_", architecture, registerFileCmd, " -o=$@ ",
      cmJoin(linkDeps, " "));
    localGen->WriteMakeRule(*this->BuildFileStream, nullptr, cubin, linkDeps,
                            { command }, false);
  }
  ...
}

Otherwise, you can just use target_compile_options or add_compile_options.

Also slightly related: the CUDA_SELECT_NVCC_ARCH_FLAGS function in the FindCUDA module (deprecated since version 3.10, which added CUDA language support into CMake).

Related CUDA docs: https://docs.nvidia.com/cuda/cuda-compiler-driver-nvcc/#options-for-steering-gpu-code-generation

Note also that since CMake v3.24, there is the native value for CUDA_ARCHITECTURES: "Compile for the architecture(s) of the host's GPU(s)."