Computer Science Technical Report Enabling Code Generation within the Sparse Polyhedral Framework

Loop transformation frameworks based on the polyhedral model use libraries such as Polylib, ISL, and Omega to represent and manipulate polyhedra and use tools like CLooG to generate loops that scan the modified polyhedra. Most of these libraries are restricted to iteration space sets and memory/array access functions with affine constraints that preclude the specification of run-time reordering transformations (i.e., inspector/executor strategies) within the existing code generation tools. Automatic generation of inspector and executor code is important for the parallelization and data locality improvements in irregular computations such as those that manipulate sparse data structures. We enable the specification of run-time reordering transformations at compile time in the Sparse Polyhedral Framework (SPF) by representing indirect memory references and run-time generated data and iteration reorderings using uninterpreted function symbols. This paper presents techniques for manipulating abstract sets and relations that include affine constraints with uninterpreted function symbols thus enabling code generation for run-time reordering transformations in the SPF.