Representation of Particle Shift in the FHP Model by TensorFlow Operations

This paper presents the tensor approach to the description of cellular automaton (CA) for modeling physical process. We consider a gas (or liquid) flow as a process for modeling. The functioning of the CA model of this process consists of two phases: collision phase and shift phase. The two-dimensional cellular-automata model of the process and approach to program implementation on supercomputers are well known. This approach is not convenient because researcher may not have special skills of parallel programming and access to a supercomputer. We offer an alternative approach for software implementation of cellular automata, which is based on the use of modern graphics adapters. Modern graphics adapters are also well-organized supercomputers, consisting of several specialized computational cores and allowing execution of operations in parallel. Compared to clusters, graphics adapters are available for a wide range of users and we believe that their capabilities are enough to implement cellular automata. To create a parallel calculation we use framework TensorFlow. The main data structure in TensorFlow is a multidimensional matrix which in terms of this framework is called a tensor. We propose to describe the global state evolution of CA by operations on tensors. As one of such operations we introduce a two dimensional bitwise convolution. This paper presents the tensor approach to the description of the shift phase in the cellular-automaton model of two-dimensional flow. Verification of the adequacy of the tensor approach is the subject of further research.