The Acceleration Of Matrix Inversion Based On Two-dimensional Mesh NoC

The development of supercomputers inspires people a lot. According to the"PC frame of reference standards","supercomputer frame of reference standards"aiming at multi-core research is proposed, namely, referring to organization structure and developing history of supercomputer to explore on multi-core system. In the supercomputer theory, the parallel-operation of program is the core problem while the core of parallel program is the parallel of algorithm.The essential of high-density computing problem is matrix operations. Based on researching the relationship between software and hardware, this paper first investigates realizing technique of big-matrix operation which could be optimized in multi-core systems. In detail, utilizing the large-dimension complex matrix operation (block, inversion, transpose, triangulation etc) as the object, the research is focusing on parallel algorithm of matrix operation, as well as parallel algorithm factorization, re-combination and task distribution course of high-density computing in the isomorphism multi-core structure. By balancing core-to-core loads of the multi-core system, the minimum routing path could be found in order to optimize resource distribution, succeeding in fully exploring the parallel capability and increase computing speed of the multi-core system with the drive of high-density operation task.The main contents are as follows:1. NoC isomorphism multi-core system is designed based on the two-dimensional grid structure.2. According to parallel algorithm theory, the existing matrix inversion algorithm is compared. At the same time, based on block LU factorization, a matrix inversion algorithm with a high parallelism is selected, as well as exploring its parallel algorithm.3. Complying with the rules, such as keeping the balance of core-to-core loads in multi-core system, decreasing the synchronous cost, and optimizing resource distribution, matrix inversion parallel algorithm is mapped into two-dimensional grid NoC multi-core structure. This could fully play the parallel capability of multi-core system and improve the computing speed of matrix inversion operation 4. The FPGA prototype on acceleration effect of matrix inversion parallel algorithm is achieved in the mapping of isomorphism multi-core. The results indicate that the method of task distribution is feasible and the acceleration effect is evident.