Abstract

A high-level architectural synthesis can be described as the process of transforming a behavioral description into a structural description. The scheduling, processor allocation, and register binding are the most important tasks in the high-level synthesis. In the past, it has been possible to focus simply on the delays of the processing units in a high-level synthesis and neglect the wire delays, since the overall delay of a digital system was dominated by the delay of the logic gates. However, with the process technology being scaled down to deep-submicron region, the global interconnect delays can no longer be neglected in VLSI designs. It is, therefore, imperative to include in high-level synthesis the delays on wires and buses used to communicate data between the processing units i.e., inter-processor communication delays. Furthermore, the way the process of register binding is performed also has an impact on the complexity of the interconnect paths required to transfer data between the processing units. Hence, the register binding can no longer ignore its effect on the wiring complexity of resulting designs. The objective of this thesis is to develop techniques for an interconnect-aware high-level synthesis. Under this common theme, this thesis has two distinct focuses. The first focus of this thesis is on developing a new high-level synthesis framework while taking the inter-processor communication delay into consideration. The second focus of this thesis is on the developing of a technique to carry out the register binding and a scheme to reduce the number of registers while taking the complexity of the interconnects into consideration. A novel scheduling and processor allocation technique taking into consideration the inter-processor communication delay is presented. In the proposed technique, the communication delay between a pair of nodes of different types is treated as a non-computing node, whereas that between a pair of nodes of the same type is taken into account by re-adjusting the firing times of the appropriate nodes of the data flow graph (DFG). Another technique for the integration of the placement process into the scheduling and processor allocation in order to determine the actual positions of the processing units in the placement space is developed. The proposed technique makes use of a hybrid library of functional units, which includes both operation-specific and reconfigurable multiple-operation functional units, to maximize the local data transfer. A technique for register binding that results in a reduced number of registers and interconnects is developed by appropriately dividing the lifetime of a token into multiple segments and then binding those having the same source and/or destination into a single register. A node regeneration scheme, in which the idle processing units are utilized to generate multiple copies of the nodes in a given DFG, is devised to reduce the number of registers and interconnects even further. The techniques and schemes developed in this thesis are applied to the synthesis of architectures for a number of benchmark DSP problems and compared with various other commonly used synthesis methods in order to assess their effectiveness. It is shown that the proposed techniques provide superior performance in terms of the iteration period, placement area, and the numbers of the processing units, registers and interconnects in the synthesized architecture