In this paper, a synthesis method of dedicated architectures is introduced. Its aim is to produce optimized systems derived from the algorithmic expression of a numerical application. The approach addresses the design of dedicated systems for applications that require high numerical computations. An efficient utilization of hardware resources is achieved through the use of vector processing with an SIMD implementation. The synthesis algorithm realizes simultaneously the design of SIMD structures and the generation of the microcode needed for implementing a software pipelining of operation of the input program.>

A multiprocessor machine with two operating modes and simple synchronization mechanisms is discussed. The operating modes are the vector single-instruction multiple-data (SIMD) mode and the parallel single program stream, multiple data stream (SPMD) mode. Synchronizations are reduced to the switching between these two modes. Mixing these two operating modes offers the programmer a comfortable programming environment: vector parallelization is performed by the compiler and the hardware, and, since SPMD parallelization requires mainly the partitioning of the entire set of data, the synchronizations are then often obvious.>

Some results of the CAPSYS method are presented. It is a new high-level synthesis method whose objective is to benefit from very long induction word (VLIW) compilation and digital signal processor (DSP) synthesis experiences to generalize the automatic processor design process. The major characteristic of this approach is to synthesize architecture in a global way. Instead of a single operative area, the synthesis considers a complete architecture. It permits one to take into account memory throughput, for instance, and, more generally, exchanges with the environment of the synthesized processor. The first part of the method, the generation of architectural models, is emphasized. This procedure furnishes a full spectrum of architectural decompositions (i.e. the number and the type of the functional units) from maximally parallel ones to fully serial ones. In each one, the synthesis algorithm creates an adequacy between the machine and the microcode. Results showing the influence of memory management on synthesized processors are given.>

The partitioning/scheduling/allocation algorithm developed for the CAPSYS method is presented. The aim of this project is to define a tool able to automatically design dedicated and embedded VLIW architectures for large and complex applications. It inherits a sizable knowledge-pool from the wider field of parallel processors, VLIW compiler design and high-level synthesis. Emphasis is placed on constraints of such a system synthesis approach and more especially on the particularity of the partitioning step. The proposed solution implements a software/hardware approach based on the list-scheduling heuristic.>

OPSILA, is a description given of a general-purpose parallel architecture with two different forms of parallelism: the well-known SIMD (single-instruction, multiple data shown), a synchronous form of parallelism; and SPMD (single program, multiple data stream), which is an asynchronous mode. It is shown that OPSILA is efficient for a wide variety of image algorithms including low and high level processing. The efficiency of OPSILA is demonstrated for the low-level algorithms, through the implementation of a set of typical operations: local (convolution), global (histogram), and geometric corrections.>

There is a continuing pressure of better processing performances in numerical signal processing. Effective utilization of LSI semiconductor technology allows the consideration of multiprocessor architectures. The problem of interconnecting the components of the architecture arises. In this paper we describe a control algorithm of the Benes interconnection network in a asynchronous multiprocessor system. A simulation study of the timeshared bus, of the omega network, of the Benes network and of the crossbar network gives a comparison of performances.

This paper describes an algorithm which allows to find out the M-output level quantizer characteristics minimizing the distortion with respect to both the Mean-Squared-Error (MSE) and Mean-Absolute-Error (MAE) criterions. This algorithm can be used with any kind of signal amplitude distributions ranging from analytical probability density functions (pdf) to experimental 'density functions. In the particular case of known pdf like uniform, Gaussian, Laplacian and Gamma densities, it gives results which agree or are better than those previously published /2,4,5,6/. The same type of algorithmic procedure may also be used for block-of-samples quantization; in this case the statistics average must be replaced by the time average. In addition, the simplicity of the proposed algorithm allows to envisage a real-time, microprocessor based, block adaptive quantizer implementation in which the quantizer parameters are periodically updated and transmitted with each data block. This technique can be used, for instance, to optimally quantize the speech coding parameters derived from the low bit rates speech compression algorithm as described in [7].