Model-based synthesis and optimization of static multi-rate image processing algorithms

Abstract

Abstract-High computational effort in modern image processing applications like medical imaging or high-resolution video processing often demands for massively parallel special purpose architectures in form of FPGAs or ASICs. However, their efficient implementation is still a challenge, as the design complexity causes exploding development times and costs. This paper presents a new design flow which permits to specify, analyze, and synthesize complex image processing algorithms. A novel buffer requirement analysis allows exploiting possible tradeoffs between required communication memory and computational logic for multi-rate applications. The derived schedule and buffer results are taken into account for resource optimized synthesis of the required hardware accelerators. Application to a multi-resolution filter shows that buffer analysis is possible in less than one second and that scheduling alternatives influence the required communication memory by up to 24% and the computational resources by up to 16%. I. INTRODUCTIONAs design complexity is becoming a major barrier for technical progress because of expensive and error-prone development, new design methodologies raising the level of abstraction are becoming increasingly popular. Simulink [1] or SystemC based high-level synthesis [2] tools for instance permit to compose complex systems by communicating blocks. However, these approaches do not allow for system-level analysis like determination of required communication buffer sizes, as the blocks can contain arbitrarily complex operations. Alternative approaches like [3], [4] are restricted to a subset of sequential languages like C. However, extraction of the contained parallelism is challenging, especially as analysis on individual statements can get computationally expensive [5].In order to address these aspects, this paper presents a novel design flow for high-level synthesis of complex multi-rate image processing applications containing up-and downsamplers. It extends existing previous work by usage of latticebased buffer analysis which considers different scheduling alternatives for multi-rate systems. As the obtained results are directly taken into account during hardware synthesis, we are able to exploit tradeoffs between required communication memory and computational logic. Furthermore, in contrast to many other approaches, analysis of the overall system does not rely on solving Integer Linear Programs (ILPs) in case of acyclic problems. Instead ILPs are only required for local analysis like actor synthesis or dependency calculation in order to assure good scaling properties of our design flow.

Scalable Analysis for Multi-Scale Dataflow Models

“…In the sub-sampled convolution filter, first the image is down-sampled, then the convolution is applied to the down-sampled image and finally the filtered image is up-sampled. A block diagram of a multi-resolution filter is provided by Keinert et al [19]. Table 1 shows the analysis results for the mentioned applications using several input image sizes.…”

A deeply pipelined and parallel architecture for denoising medical images

“…The dimensioning of intermediate memories in form of shift registers between the different blocks was done by the techniques proposed in [8] and verified by VHDL simulation. Here, in particular, the buffer on connection f 0 is very expensive since block reconstruct0 has to synchronize with the worst case datapath, which leads from the first decompose step down to the maximum depth of the filter and along all of the reconstruct stages until just after the second to last step before the data can be combined to form the final output (see Figure 2).…”

Worst-case performance analysis of synchronous dataflow scenarios

“…A number of model-based design approaches [12,16,22,19] for firm real-time parallel streaming applications are based on conservative dataflow actor models such as Synchronous Dataflow [17,23] or a slight generalisation, CycloStatic Dataflow CSDF [2]. Repetitive actors on suitably arbitrated resources can be modelled accurately and with performance guarantees, as data-driven, self-timed dataflow actors.…”