template <typename T, typename T2 = T>
void dumpArray(std::ostream &so, CudaArray<T> &arr, size_t maxDumpRows = 0,
               size_t elemsPerRow = 1) {
    int chns = sizeof(T) / sizeof(T2);
    T2 *hostArr = new T2[arr.n() * chns];
    cudaMemcpy(hostArr, arr.getBuffer(), arr.n() * sizeof(T), cudaMemcpyDeviceToHost);
    dumpHostBuffer<T2>(so, hostArr, arr.n() * sizeof(T), chns * elemsPerRow, maxDumpRows);
    delete[] hostArr;
}