A configurable SVM hardware accelerator for embedded systems

A configurable Support Vector Machine (SVM) predicting hardware accelerator applied to biological signal processing is presented in this paper. There are different types of kernel functions in SVM, and our design can realize three types of kernels (linear, polynomial, RBF). Choosing spectral energy as the input data, the input range is usually very large. So the input data should first take the logarithm in order to narrow the scope. What´s more, exponential function is also needed in RBF kernel. Therefore, a configurable CORDIC-based function generator which can calculate exponential as well as logarithmic functions is presented. And the angle and iterative formulas inside CORDIC are modified in order to expand convergence range due to the fact that original CORDIC convergence range cannot meet our demand. This modification will avoid input pre-scale and post-scale, thus save hardware resources.