This publication introduces the idea of using gpu callbacks… this is useful for large time running kernels wanting to interop with host (via pinned host mem) and the host if not wanting to
poll if mem has been modified by GPU continously would be useful if some CUDA device function (stricly speaking a function exposed for usage in kernels similar to malloc is allowed now) gets added to CUDA interrupt() where this would cause the host to recieve an interrupt from GPU which we could set a callback function via a CUDA host function cudasetgpuinterruptcallbackfunction() (of course shorter name better External Image )…
What do you think tmurray?