Error: External calls are not supported (found non-inlined call to cublasGetVersion_v2)

Question

I'm trying to use the call cublasIdamax() but I got a similar error like the title. So I write a simple code to verify the version of cublas, to avoid a version mistake in signature of function. But even this simple code result in a compilation error.

Here's my code:

__global__ void getVersion(cublasHandle_t handle, int *version){
   cublasGetVersion(handle,version);
}

int main( int argc, const char* argv[] )
{
  int  *d_version;
  int  *h_version; 
  cublasHandle_t handle;
  dim3 dimBlock( 2, 2 );
  dim3 dimGrid( 1, 1 );

  cublasCreate(&handle);

  h_version = (int *)malloc(sizeof(int*));
  cudaMalloc((void**)&d_version, sizeof(int*));

  getVersion<<<dimGrid, dimBlock>>>(handle, d_version);

  cudaMemcpy(h_version,d_version,sizeof(int),cudaMemcpyDeviceToHost);//gpu->cpu
  cout << *h_version << endl;
}

I have the following error on line 3: External calls are not supported (found non-inlined call to cublasGetVersion_v2)

What I'm doing wrong?

PS.: I looked this topic https://devtalk.nvidia.com/default/topic/500814/external-calls-are-not-supported-found-non-inlined-call-to-meminit-/ but i'm still with the problem.

Answer 1

cublasGetVersion() is a host function. You cannot call it from a __global__ function.