Molecular docking is an effective method to predict protein complexes. The optimization of molecular docking algorithm is aimed at accelerating the efficiency of molecular docking, reducing the computational cost, and making full use of computer resources. This paper mainly optimizes the semi flexible docking algorithm using three schemes:(1) The first scheme is optimized on the CPU side. (2) The second scheme, on the base of the first scheme, uses the CUFFT transplantation tool CUFFTW to provide partial GPU parallel interface for the solution. (3) The third scheme is based on the compute unified device architecture (CUDA) parallel structure, which is achieved by a pure CUDA interface and the coordinated operation of CPU and GPU. The three schemes use four different proteins to test the bound docking and unbound docking. The proteins' PDB code is 1PEE, 1B6C, 4HX3 and 2SNI, respectively. The correctness of test is verified based on the minimum root mean square deviation L_RMSD, and its value is less than 5 Å, which satisfies the CAPRI medium precision structural standard. Then, under the premise of ensuring the correctness of the results, the running speed of different proteins under different schemes is tested. Finally, under the premise of ensuring the same efficiency of different protein docking, we take 1PPE as the final object to point out the rate of docking under different schemes. Experimental results show that under the same rotation step and the correctness of the program, the final optimization effect can increase the speed of the algorithm by nearly 10 times, thus greatly improving the speed of the algorithm.