We propose a new framework to model phase retrieval problem with or without additional priori information and provide the alternating direction method of multipliers (ADMM) to solve it. This framework is simple enough and the ADMM implementation is straightforward. Though other ADMM approaches bearing the same name have proposed, the new approach differs previous arts in the flexibility of tackling with additional information and more generic sampling matrix. This advantage stems from the splitting formulation of the signal domain and measurement domain constraints, this splitting introduces the graph projection in the ADMM. The counterpart of this ADMM globally converges for convex problems, while our formulation for phase retrieval is nonconvex. By the connection of ADMM and Douglas-Rachford (DR) algorithm, we provide the local convergence guarantee of ADMM for a specific case where there is no additional information on the unknown. A smoothing version of ADMM (ADMM-S) and the error bound of reconstruction are provided for noisy case. Compared to local convergence of the recent nonconvex solvers, practically, it seems that our ADMM is much less sensitive to the initialization. Numerical results show that the new ADMM corroborates markedly improved rate of success and stability in the presence of noise for Gaussian phase retrieval, though we begin ADMM from random initializations. When applying ADMM to other practical measurement model with additional information, such as support, sparsity and TV minimization, it outperforms or refines the state-of-art solvers in reconstruction quality.