Hardware/software Partitioning Research Abstract

In embedded systems, most functionality can be implemented either by special-purpose hardware or by a software program running on a general-purpose processor. Hardware/software partitioning is concerned with deciding which functions should be implemented in hardware and which ones in software. It aims at finding an optimal trade-off between conflicting requirements on performance, power, costs, etc. Therefore, partitioning is a tough but very important problem. The main characteristics of our partitioning model are the following. We consider one software context (i.e. one general-purpose processor) and one hardware context (e.g. one FPGA). Software implementation of a component is associated with a software cost, which may be e.g. the running time of the component if implemented in software. Hardware implementation of a component is associated with a hardware cost. If two communicating components are mapped to different contexts, this incurs a communication cost. If two components are mapped to the same context, then the communication between them is neglected. The software cost of a partition is the sum of the software costs of the components that are mapped to software. Similarly, its hardware cost is the sum of the hardware costs of the components that are mapped to hardware. Its communication cost is the sum of the communication costs between those components that are mapped to different contexts. While this model is simplified and constrained in several aspects, it captures the essence of hardware/software partitioning, which lies in the simultaneous optimization of several conflicting cost factors. We have defined a formal graph-theoretic model for this problem, in which the behavioral components of the system are modeled with the vertices, and their communication with the edges of a graph, and proven its NP-completeness in the general case. We also managed to identify some efficiently solvable special cases. In particular, if communication is the only significant cost factor, then the problem can be solved optimally in polynomial time. On the other hand, if communication can be neglected, then the problem is still NP-complete, but pseudo-polynomial and approximation algorithms are known for this special case. We have developed several algorithms for solving the partitioning problem. We formulated the partitioning problem as an integer linear program (ILP) which can be solved by standard ILP solvers. This method yields the exact optimum, and works in acceptable time for graphs with up to 300 nodes. We also developed an algorithm based on branch-and-bound that can solve the partitioning problem …