在CUDA中使用指针复制类

Question

I'm trying to implement something on CUDA that involves classes which have pointers to another classes inside. I found out that I need to copy memory for all of those pointers separately and here's my approach

__host__ Production * Production::allocateToDevice() {
   Production * productionOnDevice;
   size_t productionSize = sizeof(Production);
   cudaMalloc((void **) &productionOnDevice, productionSize);
   cudaMemcpy(productionOnDevice, this, productionSize, cudaMemcpyHostToDevice);

   Vertex * boundRootOnDevice;
   cudaMalloc((void **) &boundRootOnDevice, boundRoot->getMemsize());
   cudaMemcpy(boundRootOnDevice, boundRoot, boundRoot->getMemsize(), cudaMemcpyHostToDevice);

   size_t boundRootPointerSize = sizeof(Vertex *);
   cudaMemcpy(&(productionOnDevice->boundRoot), boundRootOnDevice, boundRootPointerSize, cudaMemcpyHostToDevice);

   return productionOnDevice;
}

where

class Production {

    Vertex * boundRoot;

    public:

    Production(Vertex * root);
    __host__ Production * allocateToDevice();
    __host__ __device__ Vertex * getBoundNode();
    __host__ __device__ void execute();
    __host__ __device__ size_t getSize();

};

As you can see first I copy the Production object itself, then it's only member Vector and at the end I'm copying the pointer. Yet I'm getting access violation at penultimate line. What am I doing wrong?

Answer 1

This:

   cudaMemcpy(&(productionOnDevice->boundRoot), boundRootOnDevice, 
              boundRootPointerSize, cudaMemcpyHostToDevice);

is illegal because you have passed boundRootOnDevice (a device pointer) as the source pointer in a host-to-device transfer. This is result in the host attempting to perform pointer indirection on a device address, resulting in segfault.

Presumably you want to perform a device to device copy in this case.