Double pointer multiplication in CUDA C

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,
i am trying to multiply 2 double pointer arrays declared and saved in Pinned Memory as :

int A_d;
cudaHostAlloc((void)&A_d,N*sizeof(int *),cudaHostAllocMapped);

for (i = 0; i < N; i++){

            cudaHostAlloc((void**)&A_d[i],N*sizeof(int),cudaHostAllocMapped);
              for (j = 0; j < N; j++){
               A_d[i][j]=A[i][j];
               }

}

The gpu code is:

global void kernel(int **A_d,int **B_d,int **C_d,int N)
{
int i=threadIdx.x+blockIdx.xblockDim.x;
int j=threadIdx.y+blockIdx.yblockDim.y;
if((i<N)&&(j<N)){

int sum=0;
for(int k=0;k<N;k++)
{
sum=sum+A_d[i][k]*B_d[k][j];
}
C_d[i][j]=sum;
}
}

When i have small arrays ( N=3,N=256 ) i have correct results.But when i have bigger size like N=1024 (Nrows=1024,Ncols=1024) the result is not correct.
Any ideas?
Thank you in advance!

2 
int i=threadIdx.x+blockIdx.x*blockDim.x;
int j=threadIdx.y+blockIdx.y*blockDim.y;

I’d be willing to bet it’s how you are launching the kernel / indexing things. Once N gets large enough, you start having more than one warp get launched. Previously, only one was which is why smaller problems are correct?

Shot in the dark, but seems likely.