Looks like you misunderstand the local memory, it is shared between threads in a workgroup, and you do not want all threads writing in this array simultaneously. The idea is to get rid of array and replace it wilt just 16 uint variables: AMD stores arrays in global memory when optimization cannot coerce them into registers (and optimization is not necessarily supersmart), and uint variables are always mapped to registers (then there may be register spilling if you don't have enough of them, but at least you get a warning).