如何在CUDA中为我的指针数组分配设备内存？

Question

I have the following data structures on my host:

typedef struct point{
   int x;
   int y;
}Point;

 typedef struct pair{
     Point i;
     Point j;
     float cost;
 }Pair;

Pair* pairs[n];   // allocates an array of pointers to pair

Now, I've to copy "pairs" to the GPU. So, I declare the following pointer:

Pair **d_pair;

and allocate the memory using the following:

cudaMalloc((void**)d_pair,(sizeof(Pair)+sizeof(Pair*))*n);

Now, I copy from host to device:

cudaMempy(d_pair,pair,(sizeof(Pair)+sizeof(Pair*))*n),cudaMemcpyHostToDevice);

The kernel prototype receives d_pair as:

__global__ my_kernel(Pair* d_pair[], ... ){ 
...
}

Should the above sequence of statements work as intended? If not, what modifications I make? Basically, I want to copy Pair* pairs[n]; as such to "d_pair". How do I do this?

Answer 1

It won't work: you are sending an array of pointer, but not the objects themselves. You need to have an array (or a Vector) of Pair:

Pair pairs[n]; 

And then :

Pair *d_pair;
cudaMalloc((void**)&d_pair,sizeof(Pair)*n);
cudaMempy(d_pair,pairs,sizeof(Pair)*n,cudaMemcpyHostToDevice);

By the way, this:

cudaMempy(d_pair,pair,(sizeof(Pair)+sizeof(Pair*))*n),cudaMemcpyHostToDevice);

is non sense, you allocate space for a pointer AND a Pair. Your copy use the same (sizeof(Pair)+sizeof(Pair*))*n) expression but the array pairs is (n*sizeof(Pair*)), so you are copying undefined memory.