Edit this page

Preliminary note

polyhedral.info provides information about the polyhedral compilation community including both the latest developments as well as an overview of existing work. Heavily relying on community contributions, it provides information about software tools that use polyhedral compilation techniques, the latest publications in this area as well as community events.

Polyhedral Compilation

Polyhedral compilation encompasses the compilation techniques that rely on the representation of programs, especially those involving nested loops and arrays, thanks to parametric polyhedra FEAUTRIER1988PARAMETRICWILDE1993POLYLIBLOECHNER1999POLYLIB or Presburger relations KELLY1996OMEGAVERDOOLAEGE2010ISL, and that exploit combinatorial and geometrical optimizations on these objects to analyze and optimize the programs. Initially proposed in the context of compilers-parallelizers, it is now used for a wide range of applications, including automatic parallelization, data locality optimizations, memory management optimizations, program verification, communication optimizations, SIMDization, code generation for hardware accelerators, high-level synthesis, etc. There has been experience in using such techniques in static compilers, just-in-time compilers, as well as DSL compilers. The polyhedral research community has a strong academic background, but more and more industry users start to adapt such technologies as well.

The interest of using polyhedral representations is that they can be manipulated or optimized with algorithms whose complexity depends on their structure and not on the number of elements they represent. Furthermore, generic and compact solutions can be designed that depend on program parameters (e.g., loop bounds, tile sizes, array bounds). In a word, polyhedral techniques are the symbolic counterpart, for structured loops (but without unrolling them), of compilation techniques (such as scheduling, lifetime analysis, register allocation) designed for acyclic control-flow graphs or unstructured loops. Also, compared to optimizations that handle loops or arrays as a whole, polyhedral techniques can work at the granularity of their elements, i.e., at the granularity of a loop iteration and instance of a statement (operation), and at the granularity of an array element.

The underlying theory dates back to seminal contributions such as The Organization of Computations for Uniform Recurrence Equations by Karp, Miller, and Winograd (1968) KARP1967 and Array Expansion (1988) FEAUTRIER1988ARRAY and Dataflow Analysis of Array and Scalar References (1991) FEAUTRIER1991 by P. Feautrier, and was built on important contributions in the 80s on loop transformations (work of Lamport LAMPORT1974, Banerjee, Wolfe, Allen and Kennedy, Irigoin and Triolet IRIGOIN1988, etc.) and on systolic arrays design (work of Quinton, Rajopadhye, Delosme, etc.), before the big picture emerged. The development of optimizations tools such as PIP FEAUTRIER1988PARAMETRIC (parametric integer programming) in 1988, which enables parametric optimization on polyhedra, Polylib WILDE1993POLYLIBLOECHNER1999POLYLIB (with developments since 1987, on top of Chernikova’s algorithm), and Omega KELLY1996OMEGA around 1990 PUGH1991UNIFORM (by Bill Pugh and his group) were key contributions to make such polyhedral optimizations possible. The 1978 seminal paper on abstract interpretation Automatic Discovery of Linear Restraints Among Variables of a Program by Cousot and HalbwachsCOUSOT1978 can also be considered as one axis of foundations for polyhedral analysis, even if it gave rise to an initially-disjoint community.

Since then, in the last 20 years, the research community has extended the theory and the available polyhedral tools and libraries with Ehrhart polynomials CLAUSS1996, Barvinok’s counting algorithm, lattice optimizations, Presburger formulas manipulations, code generation algorithms. The development of freely-available tools such as ISL, CLooG, Pluto, PIPS, Fadalib, Cl@k, Gecos, to quote but a few, and the introduction of polyhedral libraries in GCC and LLVM were important steps for the dissemination of polyhedral compilation. The field of applications now covers a wide range of directions such as:

  • Program analysis

    • Dependence analysis
    • Array expansion and single assignment transformation
    • Liveness and reuse/locality analysis of array elements
    • Array region analysis and approximations
  • Scheduling theory

    • Farkas lemma and multi-dimensional scheduling
    • Automatic vectorization and parallelization of nested loops
    • Link with loop transformations (interchange, fusion, retiming, etc.)
    • Loop tiling and tiling models
  • Program verification

    • Computability and decidability
    • Program termination
    • Race detections
  • Mapping algorithms

    • Systolic array design
    • Memory mapping with memory reuse (array contraction)
    • Data mapping (HPF-like), data distribution for distributed memory
    • Communication optimizations
  • Counting and Ehrhart polynomials

    • Analysis of cache misses
    • Memory size computations
    • Iteration counting (WCET)
  • Parallelizing compiler developments

    • Hardware synthesis
    • Automatic vectorizers/parallelizers
    • Compilation for GPUs
    • Iterative compilation
    • Binary optimizations and runtime optimizations

Since 2011, the research community in polyhedral code analysis and optimizations meets annually at the International Workshop on Polyhedral Compilation Techniques (IMPACT) where last results are presented and some panel discussions are organized. IMPACT started as a workshop of the conference CGO’11 and is now a regular satellite workshop of the international conference HIPEAC. A polyhedral school has also been initiated in 2013 with a double objective: to teach polyhedral techniques and tools to newcomers (typically PhD students or researchers interested to bring such techniques to their application field) and to help polyhedral specialists discuss and establish the state-of-the-art of their field. A second edition of this school is planned.