OPENCL - print a cl_mem vector to stdout; vector type is passed as template

clVectorPrintout.cpp

// print out a vector of type T on stdout 
template <typename T>
void printClVector(cl_mem &clVector, int length, cl_command_queue &commands, int printrowlen = -1)
{
	T *tmp = new T[length];
	int err = clEnqueueReadBuffer(commands, clVector, CL_TRUE, 0, sizeof(T) * length, tmp, 0, NULL, NULL );  
	if (err != CL_SUCCESS)
	{
		printf("Error: Failed to read output array! %d\n", err);
		exit(1);
	}
	if(printrowlen < 0)	// print all as one line
	{
		for(int k = 0; k < length; k++)
		{
			cout << tmp[k] << " ";
		}
		cout << endl;
	}
	else				// print rows of "printrowlen" length
	{
		for(int k = 0; k < length/printrowlen; k++)
		{
			for(int j = 0; j < printrowlen; j++)
			{
				cout << tmp[k*printrowlen + j] << " ";				
			}
			cout << endl;
		}
		cout << endl;	
	}
	delete[] tmp;	
}

// example call:
// printClVector<float>(gpu_cluster_centroids, 2*K, commands, 2);
// where gpu_cluster_centroids holds 2K floats, and you print 2 values per line