This thesis addresses the challenges of developing multitarget code - that is to say, codes whose performance is portable across different hardware targets. We identified two key challenges : the unification of the the parallelism expression and the need to adapt the format for storing data according to the target architecture. In order to develop a multitarget version of Legolas++, a linear algebra library developed at EDF R&D, we designed MTPS (Multi-Tatget Parallel Skeleton), a library dedicated to the development of multitarget codes. MTPS allows for multitarget implementations of problems that apply the same function to all the elements of a collection. MTPS then handles the adaptation of the format for storing data according to the targeted architecture. Integrating the concepts of MTPS in Legolas++ has led to the production of a multitarget prototype of Legolas++. This prototype has allowed the development of solvers whose performances near the harware limits on different hardware architectures.
