Published on Jun 05, 2023
Two-dimensional filters are usually part of the implementation of digital image processing applications. These filters process recursive sets of instructions and require high computational speed. Optimized implementations of these filters depend on the use of Application Specific Integrated Circuits (ASICs). A system with multiple parallel processing units is a feasible design option able to achieve the required computational performance.
In this paper, a loop transformation algorithm, which allows the efficient utilization of a parallel multiprocessor system, is presented. Uniform nested loops representing the image filters and the available processors are modeled as multi-dimensional data flow graphs. A new loop structure is generated so that an arbitrary number of processors available in the system can run in parallel.
Image enhancement and edge detection are well known digital image signal processing applications that may require two-dimensional (2-D) filter-like computational solutions.
These applications usually depend on computation intensive code sections, consisting of the repetition of sequences of operations. They are also characterized by the multi-dimensionality of the data involved. An effective technique in improving the computing performance of such applications has been the design and use of Application Specific Integrated Circuits (ASICs). This paper presents a new technique applicable to the design of a 2-D filter system using multiple parallel processors. A multi-dimensional retiming algorithm embedded in this new technique provides the fully parallel utilization of the available processors, thus reducing the overall execution time of the filter function. Parallel architectures are an important tool in ASIC design.
However, these architectures require a careful partitioning of the problem in order to improve the utilization of the parallel processors.During the circuit design phase, nested loop structures can be coded using hardware description languages, such as VHDL constructs, in order to reduce the design time. However, in VHDL, the loop control indices will represent the number of times a section of the circuit will be replicated in the final synthesis under the assumption that there are no physical or cost constraints in the circuit implementation . In this paper, a multi-dimensional retiming technique is used to transform the loop in such a way to produce the parallel solution for the problem for a given number of processing units. Such a solution can then be implemented on a standard multiprocessor architecture.
