Quantitative Prediction for Early Design Space Exploration in Delft WorkBench: An Outlook

In this paper, we discuss Quipu our multidimensional quantitative prediction model for hardwaresoftware partitioning. The proposed model is based on linear regression between software metrics determined on a dataset of 127 kernels and measures from their corresponding hardware designs. These software metrics capture the complexity of the C language description. The hardware designs are determined using the DWARV C-toVHDL translator [1]. Currently, Quipu exhibits a relatively large error compared to lower level approaches, however the Quipu model can make fast and early predictions and is applicable to a wide variety of applications. For the moment, we have only considered prediction of area measures, like the number of slices or flip-flops. The main steps to improve Quipu are the following: 1) re-evaluation of the selected software metrics. 2) use of a lower level representation of the C code. 3) extension of the set of kernels. 4) extension of the modeled hardware parameters. In other words, a consolidated model can provide more, and more accurate information. We conclude that fast and early prediction of hardware characteristics is important, but our approach was not accurate enough in the past. While a somewhat larger error is acceptable in the early stages of design, we need to improve our Quipu model. Furthermore, for Quipu to be applicable, it must predict additional hardware measures for a wider range of application domains. Keywords—Reconfigurable architectures, Modeling, Estimation, Statistics, Software metrics, System analysis and design