etl VS libcudacxx

There are also countless libraries that do this. Check out ETL as an example. Notice that yea, you might have some trouble compiling this in parallel, but generally when you work with codebases over a million lines, you are compiling multiple binaries, which CAN be compiled in parallel. And since you only ever include things once, you technically don't need the pragma in your headers (though I usually put it or include guards there just to be nice). Additionally, it's less work for your preprocessor and compiler since it has fewer includes to process, and it doesn't need to deal with the forward declarations we normally paste all over the place with .h/.cpp architectures. Of course, main.cpp takes a hot minute to compile. But it also doesn't need to compile other cpp files in this binary, so it does benefit from that.

Some examples are Matrix Template Library, Blitz++, Expression Template Library, the popular Eigen and the industry standard Armadillo.

If you are thinking of using new features through https://github.com/NVIDIA/libcudacxx , you'll have to learn c++

Checkout https://github.com/NVIDIA/libcudacxx

And an implementation of the C++ standard library for CUDA.