LLVM Based Parallelization of C Programs for GPU

The paper proposes an approach to semi-automatic program parallelization in SAPFOR (System FOR Automated Parallelization). SAPFOR proposes opportunities to perform user-guided source-to-source program transformations and to reveal implicit parallelism in sequential programs. The LLVM compiler infrastructure is used to examine a program and Clang is used to perform source-to-source program transformation. This paper highlights benefits of IR-level (Intermediate Representation) program analysis which allows us to apply low-level program transformations to investigate properties of the original program. To exploit program parallelism SAPFOR relies on DVMH which is a directive-based programming model. We use subset of C-DVMH language which allows us to run parallel program on GPU as well on multiprocessors. Evaluation of presented approach has been performed using the C version of the NAS Parallel Benchmarks.