A CUDA-based parallel optimization method for SM3 hash algorithm.

Hash algorithms are among the most crucial algorithms in cryptography. The SM3 algorithm is a hash cryptographic standard of China. Because of the strong collision resistance and irreversibility of hash algorithms, they are widely used as a basic function in various fields such as digital signatures and random number generation. With the increasing real-time applications of automation in the fields of finance and office, the network puts forward higher demands for the implementing efficiency of the SM3 algorithm. We present a CUDA-based parallel optimized method for SM3 algorithm by four different ways: They are Single data stream with Single thread (SS), Multiple data streams with Single thread (MS), Single data stream with Multi-thread (SM), and Multiple data streams with Multi-thread (MM). The experimental result shows MM is the best of the four. When considering the data transmission between CPU and GPU, the proposed optimized algorithm achieves a peak performance of 166.42 Gb/s, which is 1.96 times of the best-known implementation of the SM3 algorithm on GPU platforms. Without transmission time counting, the peak performance is near 8500 Gb/s. Compared with other SM3 GPU algorithms, the algorithm proposed in this paper significantly enhances the efficiency of digest generation. Furthermore, the results show a new conclusion that the optimization of logical operations in the SM3 algorithm has reached a very high extent and the data transmission of PCIE becomes the bottleneck in the CPU+GPU data processing mode. Therefore, future work on the optimization of the SM3 algorithm should pay more attention to the PCIE data transfer efficiency. [ABSTRACT FROM AUTHOR]