Abstract
This thesis report is focused on studying data flow accelerators dedicated
to image using mathematical morphology. The main objective is to provide a
programmable and efficient implementation of basic morphological operators,
and to assemble them in such a way as to provide complex operators with
fast operation. In recent years, morphological algorithm research has been
oriented towards finding elegant algorithms to compute these complex
operators, such as watershed using priority queues. These complex
algorithms often use specific data structures that are hard to deploy on
platforms other than single-core, general-purpose processors. Moreover,
these processors continue their development in the field of parallelism by
heightening the number of cores. And because the frequency wall seems to
have been reached, the best way to optimise performance is to use
parallelising techniques. Consequently, we decided on fast implementations
of complex mathematical morphology operators, based on highly parallel
simpler operations. In the first part, we study existing computational
kernels for neighbourhood processors and suggest new ones based on recent
advances in mathematical morphology. In the second part, we use the
neighbourhood processors as building blocks to generate and manage pipeline
using high-level tools in a system on chip context. In the third part, we
present a description of a basic VLIW processor using vector instructions
deployed in a dataflow context to exploit spatial and temporal parallelism.
Finally, we analyse the performance of our system against a multi-core
workstation processor, and against a graphics processor to show the
relevance of our approach.