Optimization and acceleration of flow simulations for CFD on CPU/GPU architecture

With the increasing requirement of high computational power in computational fluid dynamics (CFD) field, the graphic processing units (GPUs) with great floating-point computing capability play more important roles. This work explores the porting of an Euler solver from central processing units (CPUs) to three different CPU/GPU heterogeneous hardware platforms using MUSCL and NND schemes, and then the computational acceleration of one-dimensional (1D) Riemann problem and two-dimensional (2D) flow past a forward-facing step is investigated. Based on hardware structures, memory models and programming methods, the working manner of heterogeneous systems was firstly introduced in this paper. Subsequently, three different heterogeneous methods employed in the current study were presented in detail, while porting all parts of the solver loop to GPU possessed the best performance among them. Several optimization strategies suitable for the solver were adopted to achieve substantial execution speedups, while using shared memory on GPU was relatively rarely reported in CFD literature. Finally, the simulation of 1D Riemann verified the reliability of the modified codes on GPU, demonstrating strong ability in capturing discontinuities of both schemes. The two cases with their 1D computational domains discretized into 10,000 cells both realized a speedup exceeding 25, compared to that executed on a single-core CPU. In simulation of the 2D step flow, we came to the highest speedups of 260 for MUSCL scheme with 800x400 mesh size and 144 for NND scheme with 400x200 computational domain, respectively.