Porting and Optimizing G-BLASTN to the ROCm-based Supercomputer

ROCm, the complete GPGPU ecosystem launched by AMD in 2015, is the first open-source software development platform for HPC/Hyperscale-class GPU computing and corresponds to NIVIDIA's CUDA platform. G-BLASTN widely used and indispensable in the field of bioinformatics which is an accelerated nucleotide comparison tool based on NCBI-BLAST. However, G-BLASTN is designed to run on CUDA platform which cannot run on ROCm. Given the popularity of ROCm and the importance of G-BLASTN, it's extremely important to porting and optimizing G-BLASTN on ROCm. This paper presented a new method of porting and optimization of G-BLASTN on ROCm-based supercomputers. First, this paper ported G-BLASTN to ROCm-based supercomputers. Based on the characteristics of the supercomputer's hardware configuration and the idea of offset collection of inductive nucleic acid sequences, this article proposed and implemented a new subject sequences H2D(host to device) transfer mode to replace the original time-consuming and inefficient transfer mode. In addition, according to the different warp specified by the two platforms, this paper adjusted the block size of the three HIP kernel functions when they were released during the word search phase. The experimental results show that performance of the optimized H2D transfer mode improves 23.6 times, and the performance of the three kernel functions improves 7.36 times, 3.02 times and 6.03 times respectively. The total performance improves 1.26 times.