"device-function-maxrregcount" message while compiling cuda code

I am trying to write a code which performs multiple vector dot product inside the kernel. I'm using cublasSdot function from cublas library to perform vector dot product. This is my code:

using namespace std;
__global__ void ker(float * a, float * c,long long result_size,int n, int m)
{
float *sum;
int id = blockIdx.x*blockDim.x+threadIdx.x;
float *out1,*out2;
int k;

if(id<result_size)
        {
                cublasHandle_t handle;
                cublasCreate(&handle);
                out1 = a + id*m;
                for(k=0;k<n;k++)
                {
                        out2 =a + k*m;
                        cublasSdot(handle, m,out1,1,out2,1,sum);
                        c[id*n + k]= *sum;
                }
        }
}
int main()
{
int n=70000,m=100;
long result_size=n;
result_size*=n;
float * dev_data,*dev_result;
float * data = new float [n*m];
float * result = new float [result_size];
for (int i = 0; i< n; i++)
        for(int j = 0; j <m;j++)
        {
           data[i*m+j]=rand();
        }

cudaMalloc ((void**)&dev_data,sizeof(float)*m*n);
cudaMalloc ((void**)&dev_result,sizeof(float)*result_size);
cudaMemcpy( dev_data, data, sizeof(float) * m* n, cudaMemcpyHostToDevice);
int block_size=1024;
int grid_size=ceil((float)result_size/(float)block_size);
ker<<<grid_size,block_size>>>(dev_data,dev_result,result_size,n,m);
cudaDeviceSynchronize();
cudaMemcpy(result, dev_result, sizeof(float)*(result_size), cudaMemcpyDeviceToHost);
return 0;
}

I have included cublas_v2 library and used the following command to compile the code:

nvcc -lcublas_device -arch=sm_35 -rdc=true  askstack.cu -o askstack

But I got the following message:

ptxas info    : 'device-function-maxrregcount' is a BETA feature

Can anyone please let me know what should I do regarding this message?

Solution

This message is informational, as said by talonmies.

This maxregcount option of NVCC is used to specify a limit of registers that can be used by a kernel and all the device functions it uses :

If a kernel is limited to a certain number of registers with the launch_bounds attribute or the --maxrregcount option, then all functions that the kernel calls must not use more than that number of registers; if they exceed the limit, then a link error will be given.

See : NVCC Doc : 6.5.1. Object Compatibility

It seems that device-function-maxregcount is used to override this value for device functions only. So, you can have a different maximum amount of registers allowed on kernels and device functions.

For device functions, this option overrides the value specified by --maxregcount.

Source : The CUDA Handbook