I got the wrong result from matrix summation

Accelerated Computing CUDA CUDA Programming and Performance
1

Hello,

I am a beginner of CUDA.
I have a problem about matrix summation(sample from CUDA_C_Programming_Guide).
Syntax is OK but I got the worng result.
Hope someone can tell me what is wrong?

###################
result:
0 , 0 , 0 ,
0 , 0 , 0 ,
0 , 0 , 0 ,
###################

OS: Linux
CUDA: V4
VGA: GeForce 210

The code is as below.
//

global void matrix_add( int a[3][3], int b[3][3], int c[3][3] ) {

int i = threadIdx.x;
int j = threadIdx.y;
c[i][j] = a[i][j] + b[i][j];

}

void matrix_2() {
int ary_size = 3;

int ary_a[3][3] = {1,2,3,4,5,6,7,8,9};
int ary_b[3][3] = {10,20,30,40,50,60,70,80,90};
int ary_c[3][3] = {0};

int ary_Da[3][3], ary_Db[3][3],ary_Dc[3][3];

int m_size = ary_size * ary_size * sizeof(int);	


cudaMalloc( (void**)ary_Da, m_size);
cudaMalloc( (void**)ary_Db, m_size);
cudaMalloc( (void**)ary_Dc, m_size);

cudaMemcpy(ary_Da,ary_a,m_size,cudaMemcpyHostToDevice);
cudaMemcpy(ary_Db,ary_b,m_size,cudaMemcpyHostToDevice);

dim3 threadPerBlock(3,3);
dim3 blockPerGrid(1,1);

matrix_add<<<blockPerGrid,threadPerBlock>>>(ary_Da,ary_Db,ary_Dc);

cudaMemcpy(ary_c,ary_Dc,m_size,cudaMemcpyDeviceToHost);

int i = 0;
int j = 0;
while(i< ary_size) {
	
	j =0;
	while(j < ary_size) {
		cout << ary_c[i][j] << " , ";
		j++;
	}
	
	cout << endl;
	i++;
} 

cudaFree( ary_Da );
cudaFree( ary_Db );
cudaFree( ary_Dc );

}

int main(int argc, char** argv)
{
matrix_2();

return 0;

}

2

Always check return codes for errors! This would have pointed you to the fact that already the cudaMalloc() calls had a mistake. They should be

cudaMalloc( (void**)&ary_Da, m_size);

cudaMalloc( (void**)&ary_Db, m_size);

cudaMalloc( (void**)&ary_Dc, m_size);

The other problem is that you allocate [font=“Courier New”]ary_Da[/font], [font=“Courier New”]ary_Db[/font] and [font=“Courier New”]ary_Dc[/font] as arrays on the host. When you use them in the kernel invocation, they then get converted to host pointers, which of course can’t work. You need to allocate them as pointers right from the beginning

int (*ary_Da)[3], (*ary_Db)[3], (*ary_Dc)[3];

Although I prefer writing

int (*ary_Da)[3][3], (*ary_Db)[3][3], (*ary_Dc)[3][3];

which then requires to rewrite the kernel as

__global__ void matrix_add( int (*a)[3][3], int (*b)[3][3], int (*c)[3][3] ) {

int i = threadIdx.x;

    int j = threadIdx.y;

    (*c)[i][j] = (*a)[i][j] + (*b)[i][j];

}
3

Hi tera,

Thanks a lot.
It’s working.

I need some reference of CUDA development.
Can you recommend me some reference with simple examples.

Regards