This paper considers downlink multiuser multiple-input-multiple-output system with parallel spatial modulation scheme, in which base station transmitter antennas are separated into K groups corresponding to K user terminals. Generalized spatial modulation is employed, in which a subset (more than single antenna) of transmit antenna array are activated and the activating pattern corresponds to specific spatial symbol. Different from existing precoding-based algorithms, we develop a two-stage detection scheme at each user terminal: In the pre-processing stage, a minimax algorithm is proposed to identify the indices of active antennas, where the key idea is that the minimum output energy of the detector is maximized; A constrained minimum output energy algorithm is proposed in the post-processing stage to mitigate multiuser interference and extract temporal symbols. Compared with existing precoding-based algorithms, the complexity is significantly reduced. Moreover, the proposed algorithm is semi-blind, in which only a small subset of channel state information is required to identify active antennas as well as eliminate multiuser interference. Simulation results demonstrate that the proposed algorithm is near-far resistant and the spectral efficiency is extensively increased compared to the conventional spatial modulation scheme.