Browse Prior Art Database

Cell Saving Normalizer for Faster Residue Prediction of the Normalized Result

IP.com Disclosure Number: IPCOM000108698D
Original Publication Date: 1992-Jun-01
Included in the Prior Art Database: 2005-Mar-22
Document File: 3 page(s) / 97K

Publishing Venue

IBM

Related People

Faget, RR: AUTHOR

Abstract

A technique is disclosed for generating residue as part of the process of normalizing the fraction portion of an IEEE floating-point result. The residue values generated are combined and used to predict the residue of the final result. By generating residue on intermediate shifts for a normalized fraction, normalizing hardware is saved in addition to a decrease in the levels of logic required to generate the residue.

This text was extracted from an ASCII text file.
This is the abbreviated version, containing approximately 52% of the total text.

Cell Saving Normalizer for Faster Residue Prediction of the Normalized Result

       A technique is disclosed for generating residue as part
of the process of normalizing the fraction portion of an IEEE
floating-point result.  The residue values generated are combined and
used to predict the residue of the final result.  By generating
residue on intermediate shifts for a normalized fraction, normalizing
hardware is saved in addition to a decrease in the levels of logic
required to generate the residue.

      Floating-point fraction results must be normalized prior to
being rounded to form a legal IEEE binary floating-point result.
Normalization is the process of shifting a fraction such that the
leading binary '1' appears as the most significant bit.  This bit is
commonly referred to as the implicit bit since it is always the
leading '1' of the fraction and is not included in the final result
but is instead 'implied' to be part of the numerical answer.  Many
cases of floating-point results have bits to the right of the least
significant bit allowed in the final fraction result.  The bit to the
right of the least significant bit allowed becomes the 'guard' bit
and all bits to the right of the 'guard' bit are logically ORed
together to form the 'sticky' bit.  These two bits are then used to
perform IEEE binary floating-point rounding.

      In our implementation, the dataflow producing the result was
checked by residue prediction.  The bits to the right of the final
least significant bit need to be measured for residue values in order
to predict the residue of the final result.

      Together, sticky...