Generation of representative input vectors for parametric designs: from low precision to high precision

Abstract

A design that can be used for different input lengths is called a parametric design. We refer to a combinational design with a large input length (e.g., 128 bits) as a large design, and we refer to a combinational design with a small input length (e.g., 20 bits) as a small design. The internal signals in a parametric design can be partitioned into precision-dependent signals (often called “data” signals) and precision-independent signals (often called “control” signals). A set of input vectors is representative if it generates all possible combinations of control signals. Generating representative input vectors is very useful for checking the correctness of a design. We present a method for efficiently generating representative input vectors for large parametric combinational designs. The method is based on an exhaustive simulation of a small design and a conversion of the representative input vectors of the small design to the large input length. The conversion is successful if the same combinations of control signals are produced in the small design and in the large design. The method was tested on a fast IEEE floating point adder design with multiple parallel paths (in: Computer Arithmetic, Proceedings, 15th IEEE Symposium on, 2001, p. 184). A coverage of 84% was achieved compared to a coverage of 53% achieved by a random-biased method.