This function shows a nice trick to manipulate warps. It can achieve better vector utilization, because threads inside a warp are executing the same SIMD instructions and we can specify how each lane actually executes.