Polyhedral Compilation: Applications, Approximations and GPU-specific Optimizations

Polyhedral compilation has been successful in analyzing, optimizing, automatically parallelizing a�ne computations for modern heterogenous target architectures. Many of the tools have been developed to automate the process of program analysis and transformations for a�ne control parts of programs including widely used open-source and production compilers such as GCC, LLVM, IBM/XL. This thesis makes contribution to the polyhedral model in three orthogonal dimensions as follows: • Applications: Applies polyhedral loop transformations on Deep learning computation kernel to demonstrate the e�ectiveness of complex loop transformations on these kernels. • Approximations: Developes two efficient algorithms to over-approximate convex polyhedra into U-TVPI polyhedra having applications in polyhedral compilation as well as automated program verification. • GPU-Specific Optimizations: Builds end-to-end fully automatic compiler framework to generate cache optimized CUDA code begining from sequential C program by using polyhedral modelling techniques.