Optimizing the Cyclic Jacobi Algorithm of Singular Value Decomposition for DLX Architecture

This project focussed on speeding up the cyclic Jacobi algorithm of the Singular Value Decomposition for DLX architecture. Using the performance monitor, it is shown that over 96% of the time is spent in a subroutine calculates Jacobi rotations. To speed up the subroutine, the improvement are performed on two aspects. From the algorithm point of view, we rst implemented the fast Jacobi matrix multiplications , which brings our program to a more mature level, and forms the algorithm structure basis for further improvements. We then implemented the fast Jacobi rotation which requiring half as many oating point operations as standard Jacobi rotation and achieved 1.2 times speed up; From the complier point of view, we optimized the assemble code by reducing the integer operations which obtained 1.7 times speed up, and explored the instruction level parallelism by loop unrolling, rescheduling and register renaming which gave us further 1.2 times speed improvement. The cache and memory traces are also performed. The results showed that cache misses does not play an important role in the performance of the algorithm. Along with the algorithm optimization , we also modiied and enhanced the DLXsim simulator by providing its "stop info" command with the ability to calculate and display the cycle counts for a particular code section, which makes the DLXsim more convenient and practical to proole a program. Bugs are found and corrected in DLXsim simulator and DLXcc complier. Optimized DLXcc complier and limitations of the DLX architecture are also addressed.