This issue asks for a short-term solution for atomics support in GPU kernels. Particularly, wrapping functions like atomicAdd via a standalone function in the GPU module would go a long way for supporting different idioms. Note that, what those functions will provide is atomic operations on non-atomic types. We have a general desire to support that and vice versa in the language. But that's a post 2.0 effort and we should not wait for that. Maybe what we learn from GPU programming can help steer the design for that feature down the road. @ronawho will capture his thoughts on the general matter soon.
Another related question in my mind is: what does/should happen with atomic types in GPU kernels as they exist today?