Abstract

An efficient way to achieve the high accuracy of the results of computations is to increase the operands capacity. The best results in terms of the problem solution rate and effectiveness of memory using can be reached in the case when a computer system provides the possibility of dynamic capacity control of processed numbers. The suggested programming system presents an effective tool for Superprecision Parallel ARiTHmetic computations (SPARTH-computations). It is developed for STARAN-like associative array processors (AAP), which is a typical example of vertical processing systems. The system is oriented to solve a large set of vector and matrix operations. Besides, the operands length may be changed dynamically during processing. From the user’s viewpoint the system represents a parallel vector processor with programmable word length called SPARTH-processor. This provides the accuracy control of computations directly during solution of the problem. The SPARTH-processor architecture is implemented within the basic AAP-architecture.

The paper describes features of the SPARTH-processor architecture and new parallel algorithms of accurate computing of dot products and polynomials, in which the automatic selection of capacity needed for exact calculations is used. The results of an estimation of the SPARTH-processor accuracy are also presented.

File
Issue
Pages
51-61