Overzealously defining __device__ functions when compiling with nvcc about fusion

[Flast]

This is not a Fusion specific issue; BOOST_GPU_ENABLED is provided by Config. You should open this issue there.

[BenByington]

I will gladly file an issue there, but I don't agree that there is no Fusion specific issue. Granted other modules probably do things wrong as well, because the whole strategy is brittle, but that doesn't change the fact that Fusion is doing things incorrectly as well.

Config is correctly determining that we are compiling a .cu file using nvcc and it is correctly populating BOOST_GPU_ENABLED with the necessary information to mark a function is being able to run on both host and gpu. Everything that Config is trying to do in regards to cuda is correct. However what Config is not doing is saying you need to mark every function as gpu enabled.

Adding BOOST_FUSION_GPU_ENABLED (e.g. __device__ __host__) to a function isn't a magical panacea to automatically make a function compile for a GPU. There are a number of things you are not allowed to do from a device function. Calling a host-only routine is one of those things you are not allowed to do. If you mark a function as BOOST_FUSION_GPU_ENABLED and then proceed to call another function that is only compiled for the host, that is an error. Just like if you decided to use certain parts of the STL, or certain disallowed c++14 constructs. Nothing Config is going to do will help if you mark a function as compilable on the GPU, and then proceed to do things that are forbidden in that context.

[Flast]

OK, but I don't know which functions are valid or invalid. So, PR is very welcome.