SEADA: An efficient methodology for optimizing mixed-precision DNNs on multi-precision spatial architectures