A number of SRAM-based field programmable gate arrays (FPGAs) allow for partial reconfiguration (PR). Partial reconfiguration can be used to maximize the resource utilization in these FPGAs. Any large design usually consists of many modular features that are never used all concurrently. An FPGA does not need to implement all these features at the same time provided that it can be reconfigured in a reasonable amount of time to implement the features that can be used simultaneously. The use of partial reconfiguration is ideal in this case, since it allows for just the features that are no longer needed to be replaced by the newly required features. Current methodologies use both external and self partial reconfiguration for this purpose. On mature multichip (MC) systems that have not made use of the PR features of their SRAM-based FPGA(s), however, these methodologies would require changes in the existing FPGA configuration protocol and/or associated hardware outside the array.

This thesis presents a novel methodology that makes PR features available to these systems for the purpose of maximizing their FPGA resources without the modifications required by the current methodologies. The proposed methodology reuses an existing data interface to send the PR data to the array and directs this data to the FPGA's internal configuration port. A prototype of this methodology is demonstrated on a commercial color space conversion (CSC) engine design using two Xilinx Virtex-II Pro FPGAs. In addition, the effectiveness of the proposed methodology is quantified by comparing the FPGA resource utilization of the original CSC engine design and that of the partial reconfigurable prototype above. Finally, since the application of partial reconfiguration inherently adds latency to the output of any design, the effects of the proposed methodology on the performance of the CSC engine are also studied and reported. This information will show that reconfiguring and loading the prototyped CSC engine in addition to processing a full image in it takes 683ms, which is within the target of one second.