| 1. |
- Thörnberg, Benny
(author)
-
Memory modeling and synthesis for real-time video processing systems
- 2006
-
Doctoral thesis (other academic/artistic)abstract
- In this thesis, a new design methodology and new tools for modeling and synthesis of real-time video processing systems are presented. A real-time video processing system is a system that performs computations on a continuous sequence of images. Image processing is a memory intensive application. This, in turn, leads to the design challenge of bridging the classical gap of speed between memories and computational units. Several techniques exist for building memory hierarchies that exploit data- locality and reuse in order to overcome this memory gap. However, the support from tools to aid the designer in dataflow analysis and memory design is very modest. Additional constructs for modeling electronic systems enable well-known sequential programming languages such as C/C++ to be used for system modeling. Ocapi and SystemC, two object-oriented specification methods are compared in a case study. In this study, SystemC is found to be the most suitable specification method for video processing systems. Most operations invoked in video processing are neighborhood oriented. For a video system designer, this spatio-temporal collection of pixels represents a natural abstraction. In addition, the same pixel neighborhood reflects data dependencies that are crucial to system synthesis. An extended SystemC modeling methodology, called IMEM is presented. IMEM can be used to capture memory transactions and stream interfaces based on the pixel neighborhood as an abstraction. Two important steps towards synthesis of video systems onto Field Programmable Gate Arrays (FPGAs) are presented. These two steps are parts of a decomposition of the complete synthesis task. Firstly, the optimal sizes and placements of all FIFO-buffers in the memory system are optimized. Bit-widths, pipelining and possible sharing of FIFO-buffers among several data flow dependencies are considered at this step. Secondly, the set of FIFO-buffers are allocated onto a set of dual-ported fined grained memories. Both synthesis steps are formally modeled using network flow techniques and linear programming. In addition, a synthesis method that can automatically transform an IMEM model of a single spatial neighborhood into a multimedia processor implementation is presented. The cache and the instruction scheduler performance are both optimized by the tool. IMEM is an application specific methodology that provides the nonhardware skilled video designer with an easy programming model and an FPGA synthesis tool. Memory usage is modeled separately from computation. This is a key feature since memory usage is accepted as being the biggest design bottleneck for video processing.
|
|
