Artificial neural networks (ANNs) are systems which are derived from the field of neuroscience and are characterized by intensive arithmetic operations. These networks display interesting features such as parallelism, classification, optimization, adaptation, generalization and associative memories. Since the McCulloch and Pitts pioneering work (McCulloch, W.S., & Pitts, W. (1943), there has been much discussion on the topic of ANNs implementation, and a huge diversity of ANNs has been designed (C. Lindsey & T. Lindblad, 1994). The benefits of using such implementations is well discussed in a paper by R. Lippmann (Richard P. Lipmann, 1984): “The great interest of building neural networks remains in the high speed processing that can be achieved through massively parallel implementation”. In another paper Clark S. Lindsey (C.S Lindsey, Th. Lindbald, 1995) posed a real dilemma of hardware implementation: “Built a general, but probably expensive system that can be reprogrammed for several kinds of tasks like CNAPS for example? Or build a specialized chip to do one thing but very quickly, like the IBM ZISC Processor”. To overcome this dilemma, most researchers agree that an ideal solution should relay the performances obtained using specific hardware implementation and the flexibility allowed by software tools and general purpose chips. Since their commercial introduction in the mid- 1980’s, and due to the advances in the development of both of the microelectronic technology and the specific CAD tools, FPGAs devices have progressed in an evolutionary and revolutionary way. The evolution process has allowed faster and bigger FPGAs, better CAD tools and better technical support. The revolution process concerns the introduction of high performances multipliers, Microprocessors and DSP functions. This has a direct incidence to FPGA implementation of ANNs and a lot of research has been carried to investigate the use of FPGAs in ANNs implementation (Amos R. Omandi & Jagath C. rajapakse, 2006). Another attractive key feature of FPGAs is their flexibility, which can be obtained at different levels: exploitation of the programmability of FPGA, dynamic reconfiguration or run time reconfiguration (RTR), (Xilinx XAPP290, 2004) and the application of the design for reuse concept (Keating, Michael; Bricaud, Pierre, 2002). However, a big disadvantage of FPGAs is the low level hardware oriented programming model needed to fully exploit the FPGA’s potential performances. High level based VHDL synthesis tools have been proposed to bridge the gap between the high level application requirements and the low level FPGA hardware but these tools are not algorithmic or application specific. Thus, special concepts need to be developed for automatic ANN implementation before using synthesis tools. In this paper, we present a high level design methodology for ANN implementation that attempts to build a bridge between the synthesis tool and the ANN design requirements. This method offers a high flexibility in the design while achieving speed/area performances constraints. The three implementation figures of the ANN based back propagation algorithm are considered. These are the off-type implementation, the on-chip global implementation and the dynamic reconfiguration choices of the ANN. To achieve our goal, a design for reuse strategy has been applied. To validate our approach, three case studies are considered using the Virtex-II and Virtex-4 FPGA devices. A comparative study is done and new conclusions are given.
LP-HLS: Automatic power-intent generation for high-level synthesis based hardware implementation flow
