// includes, system
#include<cstdlib>
#include<cstdio>
#include<iostream>
// includes, project
#include <cutil_inline.h>
#include <cuda_runtime.h>

typedef unsigned bit32_t;

int myTest();

int main()
{
 myTest();
 return 0;
}
int myTest() {
int jj, kk;
    cout<<"set host data"<<endl;
    bit32_t **h_data = (bit32_t**)malloc(sizeof(bit32_t*)*3);
    bit32_t **h_data2 = (bit32_t**)malloc(sizeof(bit32_t*)*3);
    for(jj = 0; jj<3; jj++) {
        h_data[jj] = (bit32_t*)malloc(sizeof(bit32_t)*3);
        h_data2[jj] = (bit32_t*)malloc(sizeof(bit32_t)*3);
        for(kk =0; kk<3; kk++) {
            h_data[jj][kk] = (bit32_t)kk;
            cout<<h_data[jj][kk];
        }
        cout<<endl;
    }
    bit32_t *dd_data;
    size_t pi;
    cutilSafeCall(cudaMallocPitch((void**) &dd_data, &pi, 3*sizeof(bit32_t), 3));
    cutilSafeCall(cudaMemcpy2D(dd_data, pi, h_data, 3*sizeof(bit32_t), 3*sizeof(bit32_t), 3, cudaMemcpyHostToDevice));
    cutilSafeCall(cudaMemcpy2D(h_data2, pi, dd_data, pi, pi, 3, cudaMemcpyDeviceToHost));
    cout<<"print copied back data"<<endl;
    for(jj=0; jj<3; jj++){
        for(kk = 0; kk<3; kk++)
            cout<<h_data2[jj][kk];
        cout<<endl;
        free(h_data[jj]);
		free(h_data2[jj]);
		//cudaFree(dd_data[jj]);
    }
        free(h_data);
		free(h_data2);
cudaFree(dd_data);
	return 0;
}