In this paper, we present a two-level modeling approach to performance macromodeling based on radial basis function Support Vector Machine (SVM). The two-level model consists of a feasibility model and a set of performance models. The feasibility model identifies the feasible designs that satisfy the design constraints. The performance macro-model is valid for feasible designs. We formulate the feasibility macromodeling problem as a classification problem and the performance macromodeling as a regression problem and apply SVM algorithm to build the classifier and regressors correspondingly. Our experiment shows that performance macromodels for feasible designs are much more accurate, faster to train and evaluate than those without functional or performance constraints considered.