Optimizing compilers are asked to automatically achieve good performance over an increasingly larger and heterogeneous set of architectures. Complex high-level program transformations are required to address this problem, to map the proper grain of independent computation and the proper data locality to a complex hierarchy of memory, computing and interconnection resources. The polyhedral compilation framework is one of the most powerful and flexible loop transformation system, with numerous compelling results achieved in recent years in terms of automatic program optimization (CPUs, GPUs and FPGAs). But a difficult challenge remains the deployment of those research results to larger-scale programs. Indeed, this framework uses complex mathematical algorithms that are the reason for the better program performance achieved, but which are often too time consuming for production use.
The goal of this project is to significantly improve the scalability and effectiveness of polyhedral optimizations, through the design of exact optimization methods and their associated approximation heuristics for increased scalability. We will develop novel program transformation algorithms operating under hardware resources constraints, for a variety of devices currently available on heterogeneous computing systems: for multi-core CPUs using short-vector SIMD units; for FPGAs with the help of high-level synthesis tool-chain; and for GPUs. The proposed work has the potential to significantly enhance the effectiveness of optimizing compilers thereby reducing the manual performance tuning required, with significant cost savings. The developed tools will be made publicly and freely available to the research community.
