[Development] Comparing two reals in Qt code

[Alan Alpert]

On Thu, Nov 29, 2012 at 6:53 AM, Sean Harmer <sean.harmer at kdab.com> wrote:
> On Thursday 29 November 2012 15:01:05 Dominik Holland wrote:
>> Hi all,
>>
>> last month i had some performance problems on a QML animation running on
>> low end hardware (imx233).
>> I debugged that problem and it was caused because of rounding errors
>> during the comparison of two reals.
>>
>> I fixed that problem by replacing the comparison by a qFuzzyCompare() in
>> the Qt4 declarative source code.
>> Afterwards i checked whether qt5 has the same problem... and it has
>> exactly the same problem.
>>
>> Now i pushed my patch to gerrit
>> (https://codereview.qt-project.org/#change,40655) and looked again at
>> the code
>> for any other occurrences.
>>
>> What i found is, that this happens in many places and it would be a
>> really huge effort to change the code to compare the qreals in the right
>> way.
>>
>> Is this already a known problem and do you think that it is worth to try
>> to fix the code ?
>
> Also please be aware that qFuzzyCompare() is nto the right way to compare two
> arbitrary floating point values. If one of them is zero or close to zero the
> condition it tests for becomes impossible to satisfy.
>
> Something along the lines of
>
> fuzzyCompare(float p1, float p2)
> {
>     if (qFuzzyIsNull(p1))
>         return qFuzzyIsNull(p2);
>     else if (qFuzzyIsNull(p2))
>         return false;
>     else
>         return qFuzzyCompare(p1, p2);
> }
>
> is slightly better but still not fantastic. Comparing two floats that we do
> not know a priori is actually very difficult
>
> http://randomascii.wordpress.com/2012/02/25/comparing-floating-point-
> numbers-2012-edition/
>
> I would love to see a better qFuzzyCompare() implementation in Qt but the
> existing one is used in so many places it would be a nightmare to introduce
> without unwanted side effects.

What about a second qFuzzyCompare using an absolute instead of
relative epsilon? Something like 1e-5 should give a decent answer for
values between 1e5 and -1e5, which could be used for values like on
screen positions.

By having two functions for developers to choose from we'd mitigate
the a priori knowledge problem, because some developers will know
roughly which floats they're going to be testing and they can pick the
right function for that.

--
Alan Alpert