Extracting SMP parallelism for dense linear algebra algorithms from high-level specifications