Browse Prior Art Database

Variable Precision Vector Arithmetic

IP.com Disclosure Number: IPCOM000093855D
Original Publication Date: 1966-Mar-01
Included in the Prior Art Database: 2005-Mar-06
Document File: 2 page(s) / 61K

Publishing Venue

IBM

Related People

Schlaeppi, HP: AUTHOR

Abstract

This arithmetic unit organization removes the usual dilemma in the choice of an optimum word length for machines comprising n arithmetic units :AU organized to perform vector arithmetic. The n AU's are partitioned into subsets of p AU's each. Each subset can be connected so that it functions like a single AU of p-fold width when desired.

This text was extracted from a PDF file.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 63% of the total text.

Page 1 of 2

Variable Precision Vector Arithmetic

This arithmetic unit organization removes the usual dilemma in the choice of an optimum word length for machines comprising n arithmetic units :AU organized to perform vector arithmetic. The n AU's are partitioned into subsets of p AU's each. Each subset can be connected so that it functions like a single AU of p-fold width when desired.

Only n divide p vector components can then be processed concurrently. The drawing shows the data paths required for subsets of p = 2 and p = 4. From the drawing, it is demonstrated that the gating required for changing p over the powers of 2 involves only serial data paths, even for parallel AU's.

It is thus possible to exchange arithmetic precision for vector length dynamically, i.e., under control of a program parameter p. For example, vector arithmetic 256 bits wide can concurrently process either 8 components 32 bits wide, or 4 components 64 bits wide, or 2 components 128 bits wide. A possible data flow, capable of Add, Sub & Mpy in these modes, is shown in the drawing. Consequently, the available hardware can be utilized to full advantage at all times. If n is not divisible by some of the p values provided for in the hardware, full utilization is not achievable for these p values.

The same register and adder complement is used for both fixed and floating point operation modes, the latter not shown. The characteristic of a multiple precision vector component could occupy, e.g., the leading...