Chemical Kinetics on Multi-core SIMD Architectures

Chemical kinetics modeling accounts for a significant portion of the computational time of atmospheric models. Effective application of multiple levels of heterogeneous parallelism can significantly reduce computational time, but implementation on emerging multi-core technologies can be prohibitively difficult. We introduce an approach for chemical kinetics modeling on multi-core SIMD architectures we call abstract vectors which exposes multi-layered heterogeneous parallelism. The scalar types of elements of the model state data are replaced with abstract multi-dimensional types. An implementation of the vector type and its operations can then be optimized for a given multi-core platform. This approach exposes SIMD parallelism on a fixed grid by aggregating like operations. We implemented abstract vectors in the Kinetics Preprocessor (KPP) and used our modified KPP to implement SAPRC models optimized for SSE, OpenMP, and NVIDIA CUDA.