compute same attention
but optimize memory access