Parallel implementation of image understanding tasks on gated-connection networks

Presently, data parallel massively machines can be classified in two categories according to the interpre cessor communication scheme : fixed topology machines, where PES can communicate with their nearest neighbors in the interconnection graph and dynamic interconnection machine, where a routing mechanism insures arbitrary connections among processors. Generally, fixed topology machines have fast communication primitives, but rely on a programming model that does not allow efficient algorithms. On the other hand, dynamic interconnection machines rely on a very powerful (PRAM) programming model, but exhibit poor communication time and require an expensive topology.We present in this paper a new model for SIMD massively parallel machines called associative net model. It relies on an implementation of global computations based on asynchronous communications. This work is an extension of other asynchronous communication schemes [3, 61, and it uses powerful primitives mixing communication and computations. This scheme insures a fast execution of the basic primitives. More, this model allows the use of cheap topologies (eg meshes), and can be used to program algorithms with an interesting complexity in several domains like image analysis[4].Section 2 introduces some definitions and presents the basic computing model of the machine (associative nets). Section 3 describes how asynchronous computations can be used to implement these primitives. Last, some algorithmic applications of this scheme for image analysis are presented in section 4. ASSOCIATIVE NET MODELLet P be the number of processors, P = {Pi : i E [0, P -11) the set of processors in the machine, = (P, E ) the physical interprocessor connection graph, where E represents the set of edges joining processors in the connection graph. We will assume that G is a symmetric directed graph with a constant degree D.For image analysis applications, we shall consider that E is an 8-connected 2D mesh.The basic object used for communication primitives is a mgraph, that is a sub-graph of E. More precisely, g is a mgraph if its set of nodes is P and its edges are a subset of E. Mgraphs are supposed to be directed. As the set of nodes of a mgraph is implicitly always equivalent to P , a mgraph is only represented on every processor by the subset of the edges of E that is incident to the processor in the mgraph. The set of Figure 1: Connected set example. 393 0-8186-5420-1/93 $03.00 0 1993 IEEE

show abstract

A New Massively Parallel Architecture Relying On Asynchronous Communications

Mohammadi

Dulac²,

Merigot³

1994

Proceedings. Second Euromicro Workshop on Parallel and Distributed Processing

View full text Add to dashboard Cite

This paper describes the architecture and operation of a SIMD massively parallel computer, called associative mesh, based on asynchronous communications, specifically designed for applications in the computer vision domain. Associative mesh is a reconjigurable interconnection whose basic primitives combine communications and computations. We present the associative net and outline its computational model, discuss some implementation issues, and show some application algorithms.

show abstract

Parallel implementation of image understanding tasks on gated-connection networks

Cited by 3 publications

References 8 publications

Associative nets: a graph-based parallel computing model

Associative nets: a graph-based parallel computing model

Associative meshes: A new parallel architecture for image analysis applications

A New Massively Parallel Architecture Relying On Asynchronous Communications

Contact Info

Product

Resources

About