Re: Floating Point Approximations.
From: Bob Badour <bbadour_at_pei.sympatico.ca>
Date: Thu, 29 Mar 2007 00:37:56 GMT
Message-ID: <ETDOh.16737$PV3.172688_at_ursa-nb00s0.nbnet.nb.ca>
>
> Sorry, not hip to epsilon in this context. Does the above correspond
> with 50 millionths versus 32 millionths, or thereabouts?
Date: Thu, 29 Mar 2007 00:37:56 GMT
Message-ID: <ETDOh.16737$PV3.172688_at_ursa-nb00s0.nbnet.nb.ca>
>> ... >> The epsilon would be 16 times larger for BCD floats. >> ie. 5e-7 versus 2^-25
>
> Sorry, not hip to epsilon in this context. Does the above correspond
> with 50 millionths versus 32 millionths, or thereabouts?
Nope: 0.5 millionths versus 0.0298 millionths scaled by the exponent
part of the float.
For floating point representations, epsilon is the rounding factor,
which is half the least significant digit of the mantissa. If I recall
correctly, a 32-bit BCD float has a 6 decimal digit mantissa with a 2
decimal digit exponent. A 32-bit binary float has a 24 bit mantissa with
an 8 bit exponent.