I desperately think GPU programming(or specifically CUDA) needs some language level support like coroutine/async/await to organize the data flow and the executions among different dispatched device side function calls, and more on that to have some synchronize primitives between different blocks/warps etc.
Worth noting that a GPU is essentially a hardware scheduler for large numbers of small threads that yields whenever one needs to wait for memory. They don't have a great way of changing the working set of threads.
