CUDA constant memory issue: invalid device symbol with cudaGetSymbolAddress - memory-management

Question

I am trying to set constant values on my GPU's constant memory before launching a kernel which needs these values.
My code (simplified):
__constant__ size_t con_N;
int main()
{
size_t N;
size_t* dev_N = NULL;
cudaError_t cudaStatus;
//[...]
cudaStatus = cudaGetSymbolAddress((void **)&dev_N, &con_N);
if (cudaStatus != cudaSuccess) {
cout<<"cudaGetSymbolAddress (dev_N) failed: "<<cudaGetErrorString(cudaStatus)<<endl;
}
I planned to cudaMemcpy my N to dev_N afterwards.
However, all I get at this point in the code is:
cudaGetSymbolAddress (dev_N) failed: invalid device symbol
I'm working with CUDA 6.5 so it's not a quoted symbol issue, as it is in most of the Q&A I've been checking so far.
I tried to replace con_N with con_N[1] (and remove the & before con_N in cudaGetSymbolAddress parameters): same result.
As the prototype of this function is cudaGetSymbolAddress(void **devPtr , const void* symbol ), I guessed it wanted to be given my symbol's address. However, I tried with cudaStatus = cudaGetSymbolAddress((void **)&dev_N, (const void*) con_N); and I got the same message.
I'm also getting the very same error message when I remove cudaGetSymbolAddress((void **)&dev_N, &con_N) and go directly with cudaMemcpyToSymbol(&con_N, &N, sizeof(size_t)) instead.
I'm afraid I missed something essential. Any help will be greatly appreciated.

Answer 1

The correct usage of cudaGetSymbolAddress is
cudaGetSymbolAddress((void **)&dev_N, con_N)
I'm showing this with the simple example below.
As the documentation explains, the symbol should physically reside on the device. Accordingly, using &con_N in the API call appears to be meaningless, since, being cudaGetSymbolAddress a host API, accessing the address of something residing on the device directly from host should not be possible. I'm not sure if the prototype appearing in the CUDA Runtime API document should better read as `
template<class T>
cudaError_t cudaGetSymbolAddress (void **devPtr, const T symbol)
with device symbol reference instead of device symbol address.
#include <stdio.h>
__constant__ int const_symbol;
/********************/
/* CUDA ERROR CHECK */
/********************/
#define gpuErrchk(ans) { gpuAssert((ans), __FILE__, __LINE__); }
inline void gpuAssert(cudaError_t code, const char *file, int line, bool abort=true)
{
if (code != cudaSuccess)
{
fprintf(stderr,"GPUassert: %s %s %d\n", cudaGetErrorString(code), file, line);
if (abort) exit(code);
}
}
/***************/
/* TEST KERNEL */
/***************/
__global__ void kernel() {
printf("Address of symbol from device = %p\n", &const_symbol);
}
/********/
/* MAIN */
/********/
int main()
{
const int N = 16;
int *pointer = NULL;
gpuErrchk(cudaGetSymbolAddress((void**)&pointer, const_symbol));
kernel<<<1,1>>>();
printf("Address of symbol from host = %p\n", pointer);
return 0;
}

Answer 2

In my opinion, A line of your code should be fixed like below.
cudaStatus = cudaGetSymbolAddress((void **)&dev_N, con_N);
Hope this helps you.