Automatic Parallelization of Cellular Automata for Heterogeneous Platforms

Cellular Automata (CA) simulations model atomic interactions, diffusion of materials across membranes, and spread of fire or disease and allow scientists to make predictions in situations where performing physical experiments would otherwise be infeasible or impractical. Many parallel implementations of CAs exist; however, they either do not scale in terms of memory usage or are architecture-specific solutions. In this work, we present an automated parallel code-generation framework for deterministic CA applications, ACAC, that requires no programming effort from users to generate hybrid-memory parallel codes using MPI+OpenMP, MPI+CUDA, or MPI+HIP, for CPUs, NVIDIA, or AMD GPUs, respectively. We evaluate the ACAC framework for three different CA applications with four input sizes ranging from 2.36M to 236M cells. On Xeon Phi, NVIDIA, and AMD GPUs, respectively, the automated optimized parallel code from the ACAC framework for a CA with 2.36M cells yields up to 50.9×, 544.2×, and 544.2× speedup compared to their sequential baseline running on an Intel Kaby Lake processor; 1.6×, 16.8×, and 16.8× speedup compared to a state-of-the-art specialized FPGA implementation. To our knowledge, this work is the first to provide a general framework to scale CA simulations up to 10B cells.