Automatic generation of parallel and coherent code using the YAO variational data assimilation framework

Variational data assimilation estimates key control parameters of a numerical model to minimize the misfit between model and actual observations. YAO is a code generator based on a modular graph decomposition of the model; it is particularly suited to generating adjoint codes, which is the basis for variational assimilation experiments. We present an algorithm that checks the consistency of the calculations defined by the user. We then present how the modular graph structure enables an automatic and efficient parallelization of the generated code on shared memory architectures avoiding data race conditions. We demonstrate our approach on actual geophysical applications.