Load structure in gpu shared memory

1
votes

I am working with an array of structure, and each cuda block will use the data of one structure (and only one) and do a lot of computation around it. In order for the program to work I would like to load in shared memory the structure.

I have try to use the memcpy function like that :

struct LABEL_2D{
    int a;
    float * b[MAX];
};




__shared__ struct LABEL_2D self_label;

if(threadIdx.x == 0){
memcpy(&self_label,
           label+(blockIdx.x*sizeof(struct LABEL_2D)),
           sizeof(struct LABEL_2D));
}
 __syncthreads();

But on execution I got the following error : unspecified launch failure cudaGetLastError()

I am wondering if it is possible to load a structure in shared memory.

1
cudashared-memory

1 Answers

1
votes

You are not supposed to copy the data with memcpy().

You can assign the first thread as you did, to simply initialize a shared memory variable:

struct LABEL_2D{
    int a;
    float * b[MAX];
};


__shared__ LABEL_2D self_label;

if(threadIdx.x == 0){
    slef_label = label[blockIdx.x];

}
 __syncthreads();

EDIT: Deleted other workarounds because they were practically useless.