Hello All, In chaper two of the SHARC User's Manual ("Computation Units", "IEEE Floating Point Operations" is the section heading), pg 2-3 says "Denormal operands are flushed to zeros when input to a computation unit and do not heberate an underflow exception". What's a "denormal" operand? Thanks in Advance, Rick Armstrong |
|
What's a "denormal operand"?
Started by ●January 25, 2001
Reply by ●January 25, 20012001-01-25
--- In , "Richard Armstrong" <rick@r...> wrote: > Hello All, > > In chaper two of the SHARC User's Manual ("Computation Units", "IEEE > Floating Point Operations" is the section heading), pg 2-3 says "Denormal > operands are flushed to zeros when input to a computation unit and do not > heberate an underflow exception". What's a "denormal" operand? > > Thanks in Advance, > Rick Armstrong Denormal operands are NANs (not-a-number). When you look at the IEEE floating-point format you can see there are certain bit combinations which do not make a valid floating-point number (for example 0xffffffff). These are NANs. Regards, Andor |
Reply by ●January 25, 20012001-01-25
> Date: Thu, 25 Jan 2001 00:28:19 -0500 > From: "Richard Armstrong" <> > Subject: What's a "denormal operand"? > > Hello All, > > In chaper two of the SHARC User's Manual ("Computation Units", "IEEE > Floating Point Operations" is the section heading), pg 2-3 says "Denormal > operands are flushed to zeros when input to a computation unit and do not > heberate an underflow exception". What's a "denormal" operand? > > Thanks in Advance, > Rick Armstrong > Hi Rick, A floating point number is represented as sign mantissa times radix**exponent, If mantissa has first digit non-zero, the number is considered to be normalized. When the first several digits of mantissa are zeroes, the number is considered denormal. Such numbers are used e.g. in Intel x86 FPUs to avoid sudden flushes to zero, and therefore all situated between minimum positive normalized number and maximum negative normalized number. For further reference you may search on www.docs.sun.com or docs.sun.com for Goldberg's paper. Pentium's Architecture and programming manual on http://developers.intel.com also has an introduction to IEEE floating point numbers. Regards, -- Andrew Optimized Math and Vector DSP libraries for TMS320C67x. |
|
Reply by ●January 27, 20012001-01-27
--- In adsp@y..., "Andrew V. Nesterov" <nesterov@h...>
wrote: ... > Hi Rick, > > A floating point number is represented as > > sign mantissa times radix**exponent, > > If mantissa has first digit non-zero, the number is considered to > be normalized. When the first several digits of mantissa are zeroes, > the number is considered denormal. Such numbers are used e.g. in Intel > x86 FPUs to avoid sudden flushes to zero, and therefore all situated > between minimum positive normalized number and maximum negative > normalized number. Hi Andy, thanks for correcting my mistake ;). Great to have an expert like you on board. Another interesting site on floating-point standards and arithmetic (where, among others, denormals are also discussed): http://cch.loria.fr/documentation/IEEE754/numerical_comp_guide/index.h tml Regards, Andor |