Scientific computation requires more and more performance in its algorithms. New massively parallel architectures suit well to these algorithms. They are known for offering high performance and power efficiency. Unfortunately, as parallel programming for these architectures requires a complex distribution of tasks and data, developers find difficult to implement their applications effectively. Although approaches based on source-to-source intends to provide a low learning curve for parallel programming and take advantage of architecture features to create optimized applications, programming remains difficult for neophytes. This work aims at improving performance by returning to the high-level models, specific execution data from a profiling tool enhanced by smart advices computed by an analysis engine. In order to keep the link between execution and model, the process is based on a traceability mechanism. Once the model is automatically annotated, it can be re-factored aiming better performances on the re-generated code. Hence, this work allows keeping coherence between model and code without forgetting to harness the power of parallel architectures. To illustrate and clarify key points of this approach, we provide an experimental example in GPUs context. The example uses a transformation chain from UML-MARTE models to OpenCL code.