how to create a matrix in gpu and print it on cpu? - cuda

Question

This is a code to create a matrix on gpu and print it out on cpu. Can anyone tell me where am I going wrong. Thank you.
# include <stdio.h>
__global__ void create(int **d_a){
int i = threadIdx.x;
int j = threadIdx.y;
d_a[i][j] = 1;
}
void errorCheck(){
cudaError_t error = cudaGetLastError();
if(error != cudaSuccess){
// print the CUDA error message and exit
printf("CUDA error: %s\n", cudaGetErrorString(error));
exit(-1);
}
}
# define N 5
int main(){
int **d_a, **a;
a = (int**)malloc(N * sizeof(int*));
for (int i =0; i < N; i++){
a[i] = (int*)malloc(N*sizeof(int));
}
cudaMalloc((void***)&d_a, N*sizeof(int*));
for (int i =0; i < N; i++){
cudaMalloc((void**)&d_a,N*sizeof(int));
}
errorCheck();
create <<<1, N>>>(d_a);
errorCheck();
cudaMemcpy(a, d_a, (N*N)*sizeof(int),cudaMemcpyDeviceToHost);
for (int i =0; i < N; i++ ){
for (int j = 0; j < N; j++ ){
printf("%d", a[i][j]);
}
printf("\n");
}
cudaFree(d_a);
free(a);
return 0;
}
Is there something wrong with memory allocation or memcpy ?

Answer 1

Is there something wrong with memory allocation or memcpy ?
Yes on both counts.
This:
cudaMalloc((void***)&d_a, N*sizeof(int*));
for (int i =0; i < N; i++){
cudaMalloc((void**)&d_a,N*sizeof(int));
}
would have to be done like this:
cudaMalloc((void***)&d_a, N*sizeof(int*));
for (int i=0; i < N; i++){
int *row;
cudaMalloc((void**)&row, N*sizeof(int));
cudaMemcpy(d_a+i, &row, sizeof(int*), cudaMemcpyHostToDevice);
}
And then this:
create <<<1, N>>>(d_a);
errorCheck();
cudaMemcpy(a, d_a, (N*N)*sizeof(int),cudaMemcpyDeviceToHost);
would have to be done like this:
create <<<1, dim3(N,N)>>>(d_a);
errorCheck();
for(int i=0; i<N; i++) {
int* row;
cudaMemcpy(&row, d_a+i, sizeof(int*), cudaMemcpyDeviceToHost);
cudaMemcpy(a[i], row, sizeof(int) * N, cudaMemcpyDeviceToHost);
}
[All code written in browser and not tested, use at own risk]
In short, you have decided to work with an array of pointers. This requires additional CUDA API operations because the row pointers in the GPU copy are not accessible on the host by standard assignment. You must use cudaMemcpy in every case.