[cfarm-users] cfarm26 FP inaccuracies

[Jacob Bachmeyer]

On 8/30/26 05:18, Martin Guy wrote:
> On 8/30/26 2:29 AM, Jacob Bachmeyer wrote:
> > Paranoia's page 4 (when run on cfarm26) mentions:
> >
> >     Some subexpressions appear to be calculated extra
> >     precisely with about 11 extra B-digits, i.e.
> >     roughly 3.31133 extra significant decimals.
> >     That feature is not tested further by this program.
> >
> >
>> about 11 extra bits would be consistent with the x87 80-bit format 
>> vs. standard 64-bit double...
> Well spotted! I bet that's it as the differences were microscopic.
> So it's calculating intermediate results with too much precision for 
> the IEEE. I'm not sure that's a defect!

The additional precision causes the final result to round differently 
than paranoia expects and paranoia does not distinguish between extra 
precision in intermediate calculations and outright broken rounding:  
"That feature is not tested further by this program."

In short, it gets flagged as a flaw due to limitations in the testing 
tool.  The output of paranoia is quite voluminous---I only noticed the 
note about extra precision because I was comparing it between cfarm26 
and cfarm27.


-- Jacob