Question

I would like to instantiate a class in CUDA code, that shares some of its members with other threads in the same block.

However, when trying to compile the following code, I get the error:

attribute "shared" does not apply here

(nvcc version 4.2).

class SharedSomething {

public:
    __shared__ int i; // this is not allowed
};

__global__ void run() {

    SharedSomething something;
}

What is the rationale behind that? Is there a work-around to achieve the desired behavior (shared members of a class across one block)?

Answer 1

Rost explained the rationale behind the limitation. To answer the second part of the question, a simple workaround is to have the kernel declare the shared memory, and initialize a pointer to it owned by the class, e.g. in the class constructor. Example.

class Foo 
{
public:
  __device__
  Foo(int *sPtr) : sharedPointer(sPtr, gPtr) {
    sharedPointer[threadIdx.x] = gPtr[blockIdx.x * blockDim.x + threadIdx.x];
    __syncthreads();
  }

  __device__
  void useSharedData() { printf("my data: %f\n", sharedPointer[threadIdx.x]); }

private:
  int *sharedPointer;
};

__global__ void example(int *gData) 
{
  __shared__ int sData[BLOCKDIM];

  Foo f(sData, gData);

  f.useSharedData();
}

Caveat: code written in browser, unverified, untested (and it’s a trivial example, but the concept extends to real code—I have used this technique myself).