What happens when you run a CUDA kernel?

Tracing one vector-add kernel from nvcc all the way down to the warps that execute it.