Towards an automatic generation of dense linear algebra solvers on parallel architectures