Parallelization [UPM]

• State of the art prior to project:
  • Basic technology and tools for strict independence.
• Achievements:
  • New partitioning framework proposed and implemented: parametric w.r.t. concept of independence and analysis and transformation algorithms.
  • Correctness of parallelization algorithms and methodology for applying analysis information shown.
  • Parallelization algorithms and effectiveness of global analysis extensively assessed (also on ECLPS).
  • Proposed, implemented and assessed algorithms for efficient exploitation of non-strict independence - first such parallelizer; favorably subsumes strict independence-based parallelizers.
  • Achieved substantially increased speedups (specially when little information given by user) by multiple specialization.
  • Overall compiler compares very favorably with commercial systems for imperative languages in symbolic applications.
  • Identified several avenues for improvement and special cases (e.g., integer index loops) where techniques from imperative programming can be useful.