c++ cuda: cudaMallocManaged access outside of constructor - c++

Question

I have a c++ class that uses cudaMallocManaged like so:
MyMatrix::MyMatrix(int new_rows, int new_cols, int padrr, int padcc)
{
rows = new_rows;
cols = new_cols;
padr = padrr;
padc = padcc;
cout << "allocating memory" << endl;
float *data;
cudaError_t cudaStatus = cudaMallocManaged(&data, new_rows*new_cols*sizeof(float));
if (cudaStatus != cudaSuccess){
cout << cudaStatus << endl << flush;
exit(1);
}
cudaDeviceSynchronize();
cout << "allocating memory successful:" << cudaStatus << endl;
// I CAN ACCESS DATA HERE
//data[15] = 5.5; //fine
}
MyMatrix::~MyMatrix(void)
{
cudaFree(data); // delete the data array
}
I have a header .h file too:
class MyMatrix
{
public:
MyMatrix(int new_rows, int new_cols, int padr, int padt);
~MyMatrix(void);
float *data;
int padr;
int padc;
int rows;
int cols;
}
I can access the data array fine within this constructor.
However, as soon as I try to access it (read or write) outside of it, I get terminated by signal SIGSEGV (Address boundary error). E.g.,:
MyMatrix *newmat = new MyMatrix(totalr, totalc, padr, padc);
cout << (*newmat).data[0] << endl;
Or
MyMatrix newmat = new MyMatrix(totalr, totalc, padr, padc);
cout << newmat.data[0] << endl;
How can I "persist" this pointer?

Answer 1

In your constructor you create a local variable called data:
float *data;
After allocating the cuda memory and assigning the value to the local variable data, the memory address does not seem to be stored anywhere permanently. So, your memory becomes unreachable.
You have this line newmat.data, but in the constructro you never assigned any value to a member data. You did use just the local variable with the same name.

Answer 2

Your problem is that in
MyMatrix::MyMatrix(int new_rows, int new_cols, int padrr, int padcc)
{
rows = new_rows;
cols = new_cols;
padr = padrr;
padc = padcc;
cout << "allocating memory" << endl;
float *data;
cudaError_t cudaStatus = cudaMallocManaged(&data, new_rows*new_cols*sizeof(float));
if (cudaStatus != cudaSuccess){
cout << cudaStatus << endl << flush;
exit(1);
}
cudaDeviceSynchronize();
cout << "allocating memory successful:" << cudaStatus << endl;
// I CAN ACCESS DATA HERE
//data[15] = 5.5; //fine
}
the data you use in cudaMallocManaged(&data, new_rows*new_cols*sizeof(float)) is the float *data; you declared in the line above, not the data member of your class. You just need to get rid of the local float *data; so you use the data class member like
MyMatrix::MyMatrix(int new_rows, int new_cols, int padrr, int padcc)
{
rows = new_rows;
cols = new_cols;
padr = padrr;
padc = padcc;
cout << "allocating memory" << endl;
cudaError_t cudaStatus = cudaMallocManaged(&data, new_rows*new_cols*sizeof(float));
if (cudaStatus != cudaSuccess){
cout << cudaStatus << endl << flush;
exit(1);
}
cudaDeviceSynchronize();
cout << "allocating memory successful:" << cudaStatus << endl;
// I CAN ACCESS DATA HERE
//data[15] = 5.5; //fine
}