GPU Acceleration of an Adjoint Enabled Real Gas Hypersonic Flow Solver

The objective of this Phase 1 proposal is the Graphical-Processing-Unit (GPU) acceleration of a two-dimensional, unstructured, viscous Navier-Stokes (NS) solver for aerothermodynamic applications. The subject code employs a multi-species two-temperature chemically reacting real gas model and incorporates the exact discrete tangent and adjoint systems for the real-gas system, which has been used for uncertainty quantification in past work. Our approach is based on a hybrid OpenACC-CUDA strategy, which seeks to extract maximum performance while minimizing code re-writing with an eye towards code maintainability and portability. In the Phase 1 project, the entire 2D code will be ported to V100 or A100 GPUs using OpenACC directives. This will be followed with the development of specific optimized CUDA implementations for the most critical components of the code. In particular, we will investigate CUDA implementations of the implicit line solver, the chemical source term, and a representative real-gas adjoint kernel from the 2D code. The effect of the number of species and reactions, and solver block sizes will be investigated in these component CUDA implementations. Phase 1 will also investigate alternate solver approaches with reduced memory requirements for their suitability for use on GPUs. The CUDA optimized kernels will in turn be used to further optimize corresponding portions of the simpler OpenACC implementation in order to investigate the potential for performant implementations with minimal code disruption. In Phase 2, the most promising component implementations and solution techniques will be extended to three dimensions and incorporated into a fully functional three-dimensional multi-species aerothermodynamics CFD solver optimized for GPU architectures.