Performance comparison of Hypre and NVIDIA AmgX libraries on fume modeling problem

This paper considers the task of interaction of incompressible fluid with multiple objects that was introduced in computer graphics and can be reduced to three dimensional Euler equations. We proposed an approach to estimating an acceptable error under projection stage that is equal to solving of Poisson equation with mixed type boundaries on structured grid. For this stage, we used third party libraries Hypre (CPU) and NVIDIA AmgX (GPU) as an implementation of methods for SLAE solving with sparse matrices. The testing of the proposed solver was performed on system with CPU Xeon E5-2697v3 (14 cores) and GPU Tesla K40 with different numerical methods, scene types and grid sizes. It was demonstrated that the effective performance highly depends on scene specifics, therefore the questions of library, hardware and algorithm choosing are quite ambiguous. In particular, the speed-up of GPU library over its CPU-based analogue was ranged from 20% to 273 folds.