Everything goes to a raw string

It is possible to use Strf on CUDA devices, but are naturally some limitations: you can’t write to FILE* nor std::string, nor anything that is not usable in device code. Basically, you can only write raw strings.

So, the kernel — the function marked with __global__ qualifier — needs to receive a char* ( or a pointer to another character type ) pointing to where the content is to be written.

#include <strf.hpp>

__global__ void sample(char* dest, std::size_t dest_size)
{
    strf::cstr_destination writer{dest, dest_size};
    auto print = strf::to(writer);

    print("Hello World!\n");
    print("blah blah blah", 1234, '\n');

    writer.finish();
}

There is a more complete example here. You can also take a look at the unit tests that run on CUDA device here.

Static link and CMake

To build Strf as static library for CUDA, you only need to compile the file strf/strf.cu. And your code shall define the macro STRF_CUDA_SEPARATE_COMPILATION.

Alternatively, you can use Strf CMake project. In this case, you need to turn on the STRF_CUDA_SUPPORT option:

cmake -G <generator-name> -DSTRF_CUDA_SEPARATE_COMPILATION=ON …​ <path-to-source>

As explained in another document, it is possible to integrate Strf using add_subdirectory, or exporting and importing. Both solution gives you the Strf::StrfCuda CMake target, which represents the Srtf static library for CUDA.

Unsupported features on CUDA devices

The folowing features are not supported on device code: