Section: Scientific Foundations
Compilation and Synthesis for Reconfigurable Platform
The absence of compilers is one of the major limitations for the use of reconfigurable architectures in real-life applications. Therefore, the ability to compile and optimize code on reconfigurable hardware platforms from high-level specifications is the key for a real success story and is a hot topic in the research community. We continue our research efforts to offer efficient tools with close links to architectures.
Most current programming environments for reconfigurable systems consist of separate tool flows for the software and the hardware. Processor code and configuration data for the reconfigurable processing units are handcrafted and wrapped into libraries of functions. Progress beyond current practices calls for compilers capable of generating code and configurations from a high-level general-purpose programming language. Such a compiler decides which operations go into the reconfigurable processors. Loops or frequently executed code fragments are good candidates for reconfigurable platforms. For general-purpose code, this leads to several problems: it is difficult to extract sets of operations with matching granularity at a sufficient level of parallelism; inner loops of general-purpose programs often contain excess code; i.e. code that must be run on a CPU such as exceptions, function or system calls. Efforts aimed at automatic code generation for reconfigurable architectures include    .
Another approach to programming and design of reconfigurable platform, especially for special-purpose elements, is to use techniques inspired from high-level synthesis. Here also, loops are the target of the methods: the goal is to either generate special-purpose architectures made out of arithmetic operators or to produce parallel architectures. In both cases, the output may be either efficient special-purpose hardware for computation-intensive tasks or the parameters for a reconfigurable architecture. Such approaches will eventually create a bridge between compilation techniques and hardware design.
Finally, we continue to investigate floating-point to fixed-point automatic conversion with the objective to develop an open-source tool. Multimedia and signal processing are main application fields for reconfigurable platforms. In general, these algorithms are specified using floating-point operations, but, for efficiency reasons, they have to be implemented with fixed-point operations either in software for dsp cores or as special-purpose hardware. Unfortunately, fixed-point conversion is very challenging and time-consuming, typically demanding 25 to 50% of the total design or implementation time (http://www.mathworks.com/company/newsletters/digest/may04/uwb.html ). Thus, tools are required to automate this conversion. In software implementations ( dsp , mcu ), the aim is to define the optimal fixed-point specification which minimizes the code size and the execution time for a given computation accuracy constraint. This optimization is achieved through the modification of the scaling operation location and the selection of the data word-length according to the different data-types supported by dsp s. In hardware implementations ( asic , fpga ), the complete architecture has to be defined. The efficient implementation requires to minimize the architecture size and the power consumption. Thus, the conversion process goal is to minimize the operator word-length. In the fixed-point conversion process, one of the main challenge is to evaluate the fixed-point specification accuracy. For dsp -software implementation, methodologies have been proposed  ,  ,  to achieve a floating-point to fixed-point conversion leading to an ANSI-C code with integer data types. One of the key is to closely link the compilation flow to the latest dsp features. For hardware implementation, the best results are obtained when the word-length optimization process is coupled with the high-level synthesis   .