In fact, the generation of POLYBENCH databases is not trivial, since it was built during PIPS development. At the beginning, we extracted a database from the original PIPS using the PerfectClub and SPEC95 benchmarks. At this time, the implementation of operators such as convex hull, normalization and projection was not armed with timeout mechanism, thus in case of trouble, PIPS executions lasted very long or did not even terminate correctly. For example, PIPS execution on PerfectClub's ocean.f was blocked several days on account of expensive convex hull computation, then its database was generated with many polyhedra of large size. Later, the timeout mechanism that helped reduce PIPS execution times for hard cases was implemented, and the newly generated database contained many more polyhedra but of smaller size.
Moreover, the POLYBENCH framework at this time was experimental, thus only the satisfiability operator was tested, and its databases was also used for other operators such as projection, minimization. Due to changes made during PIPS development, we cannot reproduce the original databases. To save time, we decided to use two different databases, which are the initial filtered and the new sampled ones.