Non-linear function evaluation reusing matrix-vector multipliers

This paper presents a method to extend matrix-vector multipliers to support the evaluation of non-linear functions. The proposed approach introduces non-linearity by optionally overriding the input signals of the matrix-vector multiplier. The method aims to reduce the idleness of hardware resources during computation, to maximise the reuse of arithmetic units and internal structures in existing matrix-vector multipliers, and to reduce the effort in adding additional functions. From our analysis on a case with eight non-linear functions, the proposed design consumes fewer components for addition, multiplication, division, exponentiation and logarithm than a reference design with dedicated function evaluation facilities.