Polyhedral Compilation Support for C++ Features: A Case Study with CPPTRAJ

This paper reveals challenges in migrating C++ codes to GPUs using polyhedral compiler technology. We point to instances where reasoning about C++ constructs in a polyhedral model is feasible. We describe a case study using CPPTRAJ, an analysis code for molecular dynamics trajectory data. An initial experiment applied the CUDA-CHiLL compiler to key computations in CPPTRAJ to migrate them to the GPUs of NCSA’s Blue Waters supercomputer. We found three aspects of this code made program analysis difficult: (1) STL C++ vectors; (2) structures of vectors; and, (3) iterators over these structures. We show how we can rewrite the computation to affine form suitable for CUDA-CHiLL, and also describe how to support the original C++ code in a polyhedral framework. The result of this effort yielded speedups over serial ranging from 3\(\times \) to 278\(\times \) on the six optimized kernels, and up to 100\(\times \) over serial and 10\(\times \) speedup over OpenMP.