Parallelization and acceleration of the migration algorithm

The task was to speed up the migration algorithm developed by the customer using CPU and/or GPU. The difficulty was in the peculiarities of the algorithm – it couldn’t be parallelized because of algorithm’s structure. The alternative algorithm architecturally well-suited to GPGPU computations was implemented. Since both algorithms – the original and the new one – gave the approximate solution of the differential equation, their difference to some instances proved to be too significant for further processing. As the result, the new algorithm had to be abandoned. The original algorithm was partially vectorized with the use of intrinsics and parallelized with the help of OpenMP at the higher level than that the migration algorithm had. The total speed-up on Intel Core i5 reached 5 times.