Download cuda headers on laptop without GPU - cuda

Question

I'd like to write cuda kernels on my laptop without GPU, in vscode (Ubuntu 21.04). Is it possible to just download all the headers and source files somewhere, without the need of installing the full CUDA Toolkit?
That way I can still enjoy some autocompletion, but do not have to install the entire toolkit that won't function anyway as I don't have a GPU.

Answer 1

Yes, it is possible. You can download the headers from Nvidia's GitLab repo and add the path into includePath in your VSCode project C/C++ settings the same way you would add include directory of CUDA Toolkit installation.
Below Linux example of .vscode/c_cpp_properties.json where you would replace /usr/local/cuda-11.3/include with your path to the downloaded headers. If you are targeting specific device or architecture, then definining e.g. __CUDA_ARCH__=750 (for CUDA Compute capability 7.5) will help Intellisense pick correct set of CUDA intrinsic functions that will be available on those devices.
{
"configurations": [
{
"name": "Linux",
"includePath": [
"${workspaceFolder}/**",
"/usr/local/cuda-11.3/include"
],
"defines": ["__CUDA_ARCH__=750"],
"compilerPath": "/usr/bin/g++",
"cStandard": "gnu17",
"cppStandard": "gnu++14",
"intelliSenseMode": "gcc-x64"
}
],
"version": 4
}
With the setup above, NVIDIA Nsight extension should work for syntax highlighting without CUDA Toolkit installation (not tested) with .vscode/settings.json:
{
"files.associations": {
"*.cu": "cuda-cpp",
"*.cuh": "cuda-cpp"
}
}
Or alternatively, associate your CUDA kernel source files with C++ syntax highlighting in .vscode/settings.json:
{
"files.associations": {
"*.cu": "cpp",
"*.cuh": "cpp"
}
}