This letter presents a joint convolutional neural network (JCNN)-based double dimensional (DD) hybrid precoding algorithm for adaptive fully-connected subarray structure with two-layer shared phase shifters (TPSAS) full-dimensional multiple-input multiple-output (FD-MIMO) systems. Firstly, a TPSAS structure is developed to reduce hardware complexity. Then, a novel JCNN is designed to learn predicting the vectorized hybrid precoder with the vertical phase subnetwork (VpsCNN) and horizontal phase subnetwork (HpsCNN). Furthermore, the JCNN is trained by a label and a two-stage training strategy. In the first stage, minimizing residual between hybrid precoder and the label is adopted to train the JCNN. Specially, two distinct network layers are applied to meet constraints of VpsCNN and HpsCNN, respectively, including unit modulus and custom phase layers. In the second stage, estimated channel matrices are sent into the well-trained JCNN to simultaneously obtain the vertical precoder, horizontal precoder, and digital precoder. The theoretical analyses and simulation results verify that the proposed algorithm better trades off between spectral efficiency and hardware complexity than other algorithms while enhancing energy efficiency obviously.