What is the role of double buffer in pre-fix sum using CUDA

Question

This is the naive pre-fix sum implementation

__global__ void scan(float *g_odata, float *g_idata, int n)
    {
    extern __shared__ float temp[]; // allocate