Better names for CUDA kernels

[simonbyrne]

Is your feature request related to a problem? Please describe. It's difficult to match CUDA kernel names in profiles with locations in the code:

Describe the solution you'd like You can pass name to an @cuda call (https://cuda.juliagpu.org/stable/api/compiler/#CUDA.cufunction),

Ideally it would give the broadcast expression and a file:lineno where it occurred, but it probably isn't possible (broadcasted objects don't capture it) so the next best thing might be some simple summary of the broadcast tree.

@vchuravy any ideas or suggestions?

[vchuravy]

So the idea of the mangling is explicitly that the Nvidia tools can give you better results by demangling it so that the arguments are readable. cc: @maleadt