Behaviour of __shfl_down_sync operation

My kernel for warp reduce looks like below

#include <stdio.h>
__global__ void reduce(int* var ) {

        *var += __shfl_down_sync(0xffffffff, *var, 16,32);
        *var += __shfl_down_sync(0xffffffff, *var, 8,32);
        *var += __shfl_down_sync(0xffffffff, *var, 4,32);
        *var += __shfl_down_sync(0xffffffff, *var, 2,32);
        *var += __shfl_down_sync(0xffffffff, *var, 1,32);
}

int main() {
    int c=22;
    int *dev_c;
    cudaMalloc(  (void**)&dev_c, sizeof(int) ) ;
    cudaMemcpy(dev_c,&c,sizeof(int), cudaMemcpyHostToDevice);
    reduce<<<1,512>>>(dev_c);
    cudaMemcpy( &c, dev_c, sizeof(int), cudaMemcpyDeviceToHost );
    printf( "Value is %d\n", c );
    return 0;
}

I am trying to understand how it works. I have launched it with varying blocks and threads and input as 22

Input: 22
Sno Blocks  nthreads    Warpsize    Result 
1   1        32         32          704
2   1        64         32          704
3   1        512        32          704
4   1        1024       32          1232

How does it depend on Blocks and threads
why am i getting 1232 when launching for 1,1024.
Does it depend on the warpSize in __shfl_down_sync(mask,value,delta,warpSize)
Can someone please explain how shfl_down_sync operation works for a single value rather than array.