The solver works on a row-based scheme. For each row of diagonals, it applies a reduction procedure. Each row is processed \(\log_2N -1\) times, which leads to a complete reduction of the upper and lower diagonals. Because many experiments show that the algorithm fails for the number of steps greater than 8. In that case, it is recommended to limit the number of steps to 8. The input matrix is stored as five vectors, one for each diagonal.
Since the algorithm needs random memory access in every iteration, 3 copies of the whole matrix are stored internally in the solver to allow full pipelining of the implementation.
Caution
Please note that the solver is very sensitive to zeros in any of the diagonals on input data. Due to the nature of the algorithm, any zeros on the three inner diagonals will lead to an attempt to divide-by-zero and the algorithm will fail.