Error on the result vector - c++

Question

My problem is that when I make the multiplication it only multiplies the the first row of the matrix with the first element of the vector and the next elements makes them zero so the result vector gives a wrong result.
using namespace std;
#define N 100
#define F 3
#define X 7
__global__ void matvec(int *MAT, int *VEC, int *SOL) {
int bx = blockIdx.x;
int tx = threadIdx.x;
int i = 32*bx+tx;
for (int j = 0; j < X; j++) {
SOL[i] = ((MAT[i * X + j] * VEC[j]) +SOL[i]) % 2;
}
}
int main () {
int i, j;
int MAT[N][N], VEC[N], SOL[N];
int *MAT_dev, *VEC_dev, *SOL_dev;
size_t nBytes = X * X * sizeof(int);
cout << "\t- - - - - MATRIX - - - - -\n\n";
for (i = 0; i < X; i++) {
for (j = 0; j < X; j++) {
cout << "Element [" << i << "][" << j << "]: ";
cin >> MAT[i][j];
}
}
cout << endl << endl;
for (i = 0; i < X; i++) {
for (j = 0; j < X; j++) {
cout << MAT[i][j] << " ";
if (j == (X - 1))
cout << endl;
}
}
cout << endl << endl;
cout << "\t- - - - - VECTOR - - - - -\n\n";
for (i = 0; i < X; i++) {
cout << "Element [" << i << "]: ";
cin >> VEC[i];
}
cout << endl << endl;
for (i = 0; i < X; i++) {
cout << VEC[i] << " ";
}
cout << endl << endl;
cudaMalloc((void**)&MAT_dev, nBytes);
cudaMalloc((void**)&VEC_dev, nBytes);
cudaMalloc((void**)&SOL_dev, nBytes);
cudaMemcpy(MAT_dev, MAT, nBytes, cudaMemcpyHostToDevice);
cudaMemcpy(VEC_dev, VEC, nBytes, cudaMemcpyHostToDevice);
dim3 dimBlock(X,X);
dim3 dimGrid(1,1);
matvec<<<dimGrid,dimBlock>>>(MAT_dev, VEC_dev, SOL_dev);
cudaMemcpy(SOL, SOL_dev, nBytes, cudaMemcpyDeviceToHost);
cout << "\t- - - - - RESULT - - - - -\n\n";
for (i = 0; i < X; i++)
{
cout << SOL[i] << " ";
}
cout << endl << endl;
cudaFree(MAT_dev);
cudaFree(VEC_dev);
cudaFree(SOL_dev);
system("PAUSE");
return 0;
}
Thanks for the help

Answer 1

This is because the size of MAT is much larger than it should be. Basically you need N == X, which shouldn't be a problem because both are known at compile time. Memory for a 2D array is laid out in a single contiguous block, row major for C--so in your case the first row corresponds to the first 400 (sizeof(int)*N) bytes, the second row to the second 400, etc. The length of the row is called the 'stride' However, cudaMemcpy has no idea what the stride is or which elements of MAT have been filled in, it just copies the first nBytes bytes from MAT into MAT_DEV. Since nBytes is sizeof(int)*X*X and X == 7 << N the second and subsequent rows of your matrix never get copied. Only the first 196 bytes from the MAT get copied, explaining why your second row is all zeros.