A graphics processing units (GPUs) accelerated adjoint-based optimization platform for viscous flows is proposed in this paper. This is an extension of the previously developed two-dimensional Euler solver, named ADGAR (ADjoint-GARfield), implemented to be run on GPUs. First the forward formulation for viscous flows is implemented to be run on CPU and then extended to incorporate GPU acceleration using Compute Unified Device Architecture (CUDA) kernels. The forward viscous formulation is validated against existing benchmark solutions. Then, the adjoint counterpart to the inviscid forward formulation is developed and tested on CPUs. The present GPU accelerated ADGAR flow adjoint formulation is now purely auto-differentiated (AD), unlike the previous adjoint formulation that was only hand-derived. In addition, the present implementation features an improved complex step method based mesh sensitivity computation. Significant speedup of the order of 22.5x or more is observed using ADGAR for computation on a single GPU over a single CPU core for the viscous forward formulations and around 19.5x for inviscid ADjoint formulations. These preliminary results show great potential in the GPU accelerated adjoint framework for efficient aerospace design optimization applications.
展开▼
