SIMD-vectorized Chemical Source Term Evaluation

摘要

A code generation platform for single-instruction, multiple-data (SIMD) vectorized chemical source term evaluation was developed and validated against Cantera for a wide range of chemical kinetic models. Speedups of up to 1.90-2.45 × were observed over a baseline single-instruction, multiple-thread (SIMT) parallelized code. Additionally, use of a row-major data layout was found to increase both the SIMD and SIMT performance by up to 2.00-2.38 × over a column-major data layout. Larger vector-widths were found to produce speedups up to 1.8 × as compared to a small vector-width for smaller models, e.g. GRI-Mech 3.0, but as the model size increased the performance of all vector-widths were similar. Finally, the parallel scaling efficiency was found to be slightly better for larger models, and future extensions of this work were discussed.