Fast floating point vectoring algorithms and performance evaluation

In this paper, we briefly introduce our previous work-the formalization of fast rotation-based vectorization algorithms-and show how to obtain key parameters such as window size for the implementation by extensive simulation. We also show that, if the angle selection techniques presented in this paper are used in an approximate rotation setup, such as in a Jacobi based eigenvalue decomposition (EVD) algorithm, we can profit from the fact that the average latency of the vectoring unit is significantly reduced.