Reduce Memory Consumption in Smoother

[julianlitz]

template <typename T>
class BatchedTridiagonalSolver
{
public:
    BatchedTridiagonalSolver(int matrix_dimension, int batch_count, bool is_cyclic = true)
        : matrix_dimension_(matrix_dimension)
        , batch_count_(batch_count)
        , main_diagonal_("BatchedTridiagonalSolver::main_diagonal", matrix_dimension * batch_count)
        , sub_diagonal_("BatchedTridiagonalSolver::sub_diagonal", matrix_dimension * batch_count)
        , buffer_("BatchedTridiagonalSolver::buffer", is_cyclic ? matrix_dimension * batch_count : 0)
        , gamma_("BatchedTridiagonalSolver::gamma", is_cyclic ? batch_count : 0)
        , is_cyclic_(is_cyclic)
        , is_factorized_(false)
    {
        assign(main_diagonal_, T(0));
        assign(sub_diagonal_, T(0));
    }

We dont need to allocate buffer here. It can be supplied externally in the solve() method.

This reduces the memory usage in the smoother by 50% and ~10% overall.

Note that often the radial section does contain more nodes then the circular section, so this is probably less dramatic and not really noticeable.